二元论

Dualisms

在每种文化中，我们都会发现，人们会用“只有生物才有灵魂”的说法来区分非生物和生物。例如，心理学名称的来源——古希腊词“psyche”（灵魂）——意为生命的气息。据说，当一位战士死去时，他的灵魂便离开了肉体。“psyche”并非不朽，也不具备任何精神。它既不负责死者，也不负责认知功能。希腊人将这些功能赋予了位于身体其他部位（例如眼睛（感知）或心脏（情感））的大约十几个微型灵魂。由于该理论认为生物是由身体和灵魂两部分构成的，因此被称为二元论。

In every culture, we find that people separate nonliving from living things by saying that only living things have souls. For example, the ancient Greek word psyche, from which psychology takes its name, meant the breath of life. When a warrior died, it was said that his psyche had left his body. Psyche was not immortal and did not carry the personality of the deceased, nor did it perform cognitive functions, which the Greeks assigned to about a dozen mini-souls located in other parts of the body such as the eye (perception) or the heart (emotion). Because it says living beings are made of two things, body and soul, this idea is called dualism.

灵魂作为人整个人格的承载者，承担着道德责任，并能在肉体死亡后继续存在，这一概念似乎最早出现在印度教思想中，并且在除佛教之外的所有世界主要宗教中都有体现。佛教是唯一的例外，它否认灵魂的存在。意识被承认，但被视为一个过程，而非独立于身体之外的事物。由于它与科学唯物主义相容，许多认知科学家对佛教持赞赏态度，并实践佛教正念（Blackmore，2005a，b）。

Soul as a bearer of a person’s entire personality, morally responsible, and capable of surviving bodily death seems to have appeared in Hindu thought and is found in all but one of the great world religions. The exception is Buddhism, which denies the existence of souls. Consciousness is acknowledged but is regarded as a process rather than a thing separate from the body. Because it is compatible with scientific materialism, many cognitive scientists regard Buddhism favorably and practice Buddhist mindfulness (Blackmore, 2005a, b).

民间心理学，或心理理论

Folk Psychology, or Theory of Mind

除了宗教对心灵或灵魂的理解之外，还有另一种普遍的人类心理学，它不包含形而上学的信仰或关于身体位置的主张。它被称为常识心理学（社会心理学）、民间心理学（哲学）或心智理论（发展心理学）。它非常强大，使我们能够毫不费力地解释行为，是人类重要的适应性特征。

Alongside religious conceptions of the mind, or soul, there is another universal human psychology, without metaphysical commitments or claims about bodily location. It is called common sense psychology (in social psychology), folk psychology (in philosophy), or Theory of Mind (TOM) (in developmental psychology). It’s powerful, giving us the ability to explain behavior with little effort and is an important human adaptation.

请看下面的视频。你看到一个孩子在厨房里，厨房里有两个柜子，一个在左边，一个在右边。孩子从左边的柜子里拿出一个罐子，拿了一些糖果，把罐子放回左边的柜子里，然后离开。另一个孩子进来，从罐子里拿出糖果，放在右边的柜子里。第一个孩子回来后，视频停止，你被问到他会打开哪个柜子去拿糖果罐。成人和四岁以上的儿童会说是左边的柜子，而年龄较小的孩子则说回来的孩子会尝试打开右边的柜子，因为糖果罐实际上就在那里。

Consider the following video. You see a child in a kitchen with two cabinets, one on the left and one on the right. The child gets a jar from the left cabinet, takes some candy, replaces the jar in the left cabinet, and leaves. Another child enters, takes candy from the jar, and places it in the right cabinet. The first child returns, the video is stopped, and you are asked which cabinet the child will open to fetch the candy jar. Adults and children over about age four say that it will be the left cabinet, while younger children say the returning child will try the right cabinet, where the candy jar really is.

在这项错误信念测试¹（Bretherton & Beeghly，1982；Wellman，1985；Slaughter，2015）中，年纪较大的参与者表明他们知道其他人有思想，并能将错误的信念归咎于他们。心智理论假设两种心理状态相互作用导致行为。第一种状态是欲望，它为行动提供动机；在这种情况下，它是饥饿或对甜食的享受。另一种状态是信念，即关于世界如何的想法。非常年幼的孩子不会将内心的信念归因于行动者，他们会说回来的孩子会在第二个孩子给的新位置寻找糖果罐。年龄较大的儿童和成人会将一种关于糖果罐最后已知位置的私人心理状态归因于第一个孩子，并预测孩子会根据这种错误信念采取行动。

In this false belief test¹ (Bretherton & Beeghly, 1982; Wellman, 1985; Slaughter, 2015), older participants show they know others have minds and can ascribe false beliefs to them. Theory of Mind posits two mental states that interact to cause behavior. The first state is desire, which provides motivation for action; in this case, it is hunger or enjoyment of sweets. The other state is belief, an idea about how the world is. Very young children do not attribute inner beliefs to actors, saying that the returning child will look for the candy jar in the new location given to it by the second child. Older children and adults attribute to the first child a private mental state about the last known location of the candy jar and predict that the child will act upon that mistaken belief.

假设私人心灵的存在，就提出了知识的问题，也就是认知科学的起源。我们可以这样总结这个问题：我们想要知识关于世间万物的认知，而信仰是精神的。因此，真理在于事物与信仰之间的对应。

Positing the existence of private minds raises the problem of knowledge, the origin of cognitive science. We can summarize the problem this way: We want knowledge about things in the world, but beliefs are mental. Truth, therefore, lies in correspondence between things and beliefs.

知识（现实）与观点（外表）

Knowledge (Reality) vs. Opinion (Appearances)

知识问题如同硬币，有两面。一面是哲学层面的，由哲学的一个分支——认识论——来探究：什么是知识？知识（即真理）与未经证实的知识候选对象（即观点）之间的区别是什么？硬币的另一面是心理学层面的：心灵通过哪些心理过程获得观念并确定哪些观念是真实的？苏格拉底对这个哲学问题提出了强有力的、具有历史意义的论断，但他对心理学问题则留下了开放的余地，有待进一步探讨。

Like a coin, the problem of knowledge has two sides. One side is philosophical, investigated by the branch of philosophy called epistemology: What is knowledge? What distinguishes knowledge, i.e., truth, from unproven candidates for knowledge, i.e., opinion? The other side of the coin is psychological: By what mental processes does the mind obtain ideas and determine which are true? Socrates made a strong and historically decisive claim on the philosophical question, but he left the psychological question open for development.

苏格拉底的对话者认为，如果一个想法与某事物相对应，那么它就是正确的；因此，如果我说我家有 17 扇窗户，那么这个说法就是正确的，也就是知识，当且仅当我家有 17 扇窗户。然而，苏格拉底拒绝将与事物相对应作为知识的定义。假设我伸手到裤兜里掏出一些硬币，握在拳头里，问你我有多少找零。现在，你可能会猜 77 美分，而且你是对的（三个 25 美分 + 两个 1 美分）。你的猜测（想法）与我手中的找零（事物）的总和相对应，但苏格拉底会说，而且你也可能会同意，你的猜测只是偶然正确。对苏格拉底来说，这意味着一个想法要算作知识，它需要的不仅仅是与世界的对应。

Socrates’ interlocutors suggested that an idea is true if it corresponds to some thing; thus, if I claim I have 17 windows at home, that claim is true, i. e., is knowledge, if and only if my house has 17 windows. Socrates, however, rejected correspondence to things as the definition of knowledge. Suppose I reach into my pants pocket and dig out some coins, holding them in my fist and asking you how much change I have. Now, you might guess 77 cents, and you would be right (three quarters + two pennies). Your guess (idea) corresponds to the sum of the change (things) in my hand but Socrates would say, and you would likely agree, that your guess was only accidentally true. To Socrates, this meant that for an idea to count as knowledge it needs more than just correspondence to the world.

为了说明这一点，苏格拉底做出了一个典型的举动，指向专业知识。如果你想了解美食，你会向谁寻求？一位伟大的厨师。如果你想了解美，你会向谁寻求？一位艺术鉴赏家。如果你想了解正义，你会向谁寻求？一位伟大的法学家。因为他们擅长烹饪、收藏和判断，所以这些专家一定具备烹饪、美和正义的知识，因此应该能够解释善、美和正义。

To show what this is, Socrates makes a characteristic move, pointing to expertise. If you want knowledge of good food, whom do you seek? A great chef. If you seek knowledge of beauty, whom do you seek? An art connoisseur. If you seek knowledge of justice, whom do you seek? A great jurist. Because they cook well, collect well, and judge well, these experts must possess knowledge of cookery, beauty, justice, and should therefore be able to explain the good, the beautiful, and the just.

苏格拉底认识到，厨师、鉴赏家或法官有时难以清晰地表达他们的知识，他贬低这些“专家”只掌握了部分真理。他们做得不错，但却无法向他人证明自己的行为是合理的。重要的是，他不接受正确的行动可能源于直觉，而是坚持认为专家判断依赖于知识，即使这种知识是隐性的，就像在一位无法解释自身行为或决定的专家。他著名的苏格拉底式方法旨在通过一系列深入的问题，引出潜意识的知识。因此，苏格拉底并非将自己描述为真理的教师，而是真理的助产士。

Socrates recognized that sometimes a chef, connoisseur, or judge might have trouble articulating their knowledge, and he disparaged such “experts” as possessing halfway truth. They acted well but could not justify their actions to others. Importantly, he did not accept that right action might flow from intuition but insisted instead that expert judgment depended upon knowledge, even when that knowledge was tacit, as it would be in an expert who could not explain his actions or decisions. His famous Socratic method was designed to draw out unconscious knowledge through a series of penetrating questions. Socrates described himself therefore not as a teacher of truth but as a midwife of truth.

因此，对于苏格拉底来说，知识不是真正的信念，而是明确的、合理的真正信念。

For Socrates, then, knowledge was not true belief but explicit and justifiable true belief.

知识的对象是什么？

What Is the Object of Knowledge?

苏格拉底认识论的另一个重要特征对研究高级认知过程具有重要意义。让我们回到我口袋里的零钱。当你检查我的25美分和1美分硬币时，仔细观察其中一枚25美分硬币，发现它不太对劲；它有点太亮、太轻、太薄，而且一侧还有高顶礼帽的徽章。我似乎在口袋里放了魔术师称之为“掌上硬币”的东西，它和真硬币大小相同，但更容易操纵。乍一看，这似乎只是表象（乍一看像25美分）和现实（哎呀！这不是25美分；别想花掉它！）之间差距的简单例证。你一开始形成了一个无法证实的错误感知信念。

Another important feature of Socrates’ epistemology has important implications for studying the higher cognitive processes. Let’s go back to the change in my pocket. As you inspect my quarters and pennies, a closer look at one of the quarters shows it’s not quite right; it’s a little too shiny, too light, and too thin, and it has a top hat insignia on one side. I seem to have left in my pocket what magicians call a palming coin, made in the same size as a real coin, but easier to manipulate. At first glance, this appears to be a simple illustration of the gap between appearances (it looks like a quarter at first) and reality (oops! It’s not a quarter; don’t try to spend it!). You at first formed a false perceptual belief that could not be justified.

然而，这其中还有更多不为人知的奥秘。苏格拉底可能会问，为什么这枚手掌硬币不是真正的25美分硬币？你可能会说，因为它用了错误的金属，雕刻也错误。但假设我用和美国铸币局一样的白铜合金制造我的手掌硬币，并使用先进的3D打印机制造，使其在各个方面都与美国铸币局的25美分硬币难以区分，材质和外观上没有任何区别，那又如何呢？财政部仍然不会将其视为合法的、可支配的货币，而且它也没有像区块链这样的加密货币的保障。

However, there’s more here than your fooled perception. Why, Socrates might ask, is the palming coin not a real quarter? You would probably say it’s because it’s made of the wrong metal and carries the wrong engravings. But suppose that I made my palming coins out of the same cupro-nickel alloy as the US Mint, and used an advanced 3-D printer to manufacture the coin so that in every physical aspect it was indistinguishable from a US Mint quarter, there being no differences whatsoever in material or appearances? The Treasury Department would still not regard it as legitimate, spendable money, and it’s not guaranteed by blockchain-like cryptocurrencies.

苏格拉底认为，知识有两种类型，分别对应不同的认知对象。一种是感知性知识，主要针对具体事物，例如我口袋里的硬币或视频中的猫。另一种是有关对象所属类别的知识，例如 QUARTER、COIN 或 CAT。哲学中，这些对象和类别的名称是个体（或特有的）和普遍的。如今，哲学家倾向于使用token来表示个体，使用type来表示普遍的。因此，每个实体的 25 美分硬币都是 QUARTER 和 COIN 类型的 token，而每只猫都是 CAT 类型的 token。

Socrates saw that there are two types of knowledge having different objects to be known. There’s the largely perceptual knowledge of particular things, such as the coins in my pocket or a cat in a video. On the other hand, there is knowledge of the classes to which objects belong, such as QUARTER, COIN, or CAT. The names used in philosophy for these objects and classes are individuals (or particulars) and universals. Today, philosophers tend to use the terms token for individuals and type for universals. Thus, every physical quarter is a token of the types QUARTER and COIN, and every cat is a token of the type CAT.

苏格拉底说，最完美的知识是关于普遍性的知识。科学定律并非只适用于少数对象。牛顿将他的万有引力定律称为普遍引力定律，适用于宇宙中任何质量的任何两个物体。然而，普遍性的存在使得知识必须被真信念所证实的要求变得复杂。苏格拉底在处理专业知识时看到了这个问题。一位鉴赏家说，一座希腊的库罗斯（年轻人的雕像）很美，莫奈的睡莲很美，或者一次高难度的达阵接球非常漂亮。它们有什么共同点，使它们成为“美丽”这一通用类别的成员？

The most perfect knowledge, said Socrates, is knowledge of universals. The laws of science do not apply only to a few objects. Newton called his law of gravity the universal law of gravity, applying to any two objects of any mass anywhere in the universe. However, the existence of universals complicates the requirement that knowledge be justified true belief. Socrates saw the problem when dealing with expertise. A connoisseur says that a Greek kuros (statue of a young man) is beautiful, that Monet’s water lilies are beautiful, or a difficult touchdown catch was beautiful. What do they have in common that makes them members of the universal category BEAUTY?

普遍性问题对于认知科学本身及其应用都至关重要。动物是否拥有普遍的概念，还是它们只对即时刺激做出反应？儿童拥有学习语言的动力，语言中蕴含着丰富的普遍术语，他们几乎无需任何指导就能掌握。这种动力来自何处？儿童如何从几个月的特定经历中形成普遍的观念？普遍性是否局限于特定文化？还是像许多人认为的人权一样，它们真正具有人类的普遍性？如果人工智能 (AI) 真的能够像人类一样思考和决策，我们如何对它们进行编程，使它们能够理解我们自身往往难以理解的东西？它们能学会成为专家吗？或者成为有道德的战士吗？

The problem of universals is important for cognitive science both as science and as application. Do animals have universal concepts or do they respond only to immediate stimuli? Children possess a drive to acquire language, which is rich in universal terms, and they pick it up with minimal teaching. Where does the drive come from and how do children arrive at their universal ideas from a few months of particular experiences? Are universals parochial to cultures or are they truly universally human, as many people believe human rights are? If artificial intelligence (AI) is to truly be able to think and decide as humans do, how do we program them to know what we often only dimly grasp ourselves; can they learn to be experts? Or ethical soldiers?

柏拉图对这个问题的概括

Plato’s Master Formulation of the Problem

苏格拉底给认知科学留下了诸多难题。他的学生柏拉图（约公元前429-347年）将苏格拉底的洞见、传统希腊关于灵魂的观念，以及刚刚渗入希腊世界的印度教宗教观念融合在一起，形成了一个强大的框架，用于思考心灵、身体和知识，并延续了数千年。

Socrates bequeathed a tangle of problems to cognitive science. His student Plato (c. 429–347 BCE) melded Socratic insights, traditional Greek ideas about souls, and Hindu religious notions just infiltrating the Greek world into a powerful framework for thinking about mind, body, and knowledge that endured for millennia.

很少有认知科学家读过柏拉图，但他的思想却深深植根于我们的思想，以至于在不知不觉中影响着科学。我们研习古典学的意义，就像佩恩和泰勒揭示了一个伟大的魔术：有时，了解秘密会让魔术更加令人印象深刻。

Few cognitive scientists have read Plato, but his ideas are so deeply embedded in our thought that they influence science unconsciously. The point of our exercise in the classics is like Penn and Teller revealing a great magic trick: Sometimes knowing the secret makes the trick more impressive.

柏拉图的认识论

Plato’s Epistemology

柏拉图深化并整理了其老师关于认知挑战、表象与现实之间的差距以及个体与普遍性之间难以捉摸的关系的洞见。在他最具影响力的著作《理想国》中，柏拉图提出了几个隐喻来阐述他的知识理论。其中之一就是“线的隐喻”（图1.1）。

Plato sharpened and codified his teacher’s insights about the challenges of cognition, the gap between appearance and reality and the elusive relationship between individuals and universals. In his most influential work, The Republic, Plato proposed several metaphors to display his theory of knowledge. One of them is the Metaphor of the Line (Figure 1.1).

苏格拉底在户外授课时，在沙地上画了一条垂直线，标出感知或已知的事物（线左侧）与表征这些事物的心理状态（线右侧）之间的区别。然后，他用一条水平线将垂直线一分为二，水平线下方是感知到的事物及其心理表征，上方是已知事物及其心理表征。

Teaching outdoors, Socrates drew a vertical line in the sand to mark the difference between objects perceived or known, on the left side of the line, from the corresponding mental states that represent them to us, on the right side of the line. He then bisected the vertical line with a horizontal that marked things perceived and their mental representations, below the line, from things known and their mental representations, above it.

最贫乏的知识对象是图像，比如水中的倒影。一位爱开玩笑的人曾经说过，护照能让你到处旅行而不被人认出来。一个相关的问题是多对一问题，即个体与普遍性之间的区别。即使在我们一致认同的照片中，你也会因为年龄、体重、发型或面部表情的变化而显得不同。你的照片版本有很多，但你只有一个，我们想了解你，而不是你的照片。Instagram 和 Zoom 滤镜会撒谎。

The poorest objects of knowledge are images, such as reflections in water. A wag once said that a passport lets you travel everywhere unrecognized. A related problem is the many-to-one problem, the distinction between individuals and universals. You look different even in photos that we agree really look like you, across changes in age, weight, hairstyle, or facial expression. There are many image versions of you, but only one you, and we want to know you, not your images. Instagram and Zoom filters lie.

看到一个人或一个物体，就像苏格拉底知识阶梯上的一个阶梯，它引发的是信念而非想象。但感觉会产生知觉，即对世界物体的心理图像，而感觉是一种不可信的身体过程。想象一下，在一个冬夜，两个人进入一个有暖气的房间。A 一直在楼下给煤炉加煤，炉子的温度是 100 华氏度（约 37 摄氏度）。B 从外面进来，炉子的温度是 0 华氏度（约 0 摄氏度）。A 会觉得房间凉爽宜人，而 B 会觉得房间温暖宜人。但究竟是凉爽还是温暖呢？A 和 B 由于身体的物理状态不同，对房间的感知也不同：每个人都知道自己的私人体验；但他们都无法感知房间的真实温度。由于经验取决于我们（不可靠的物理）身体的状态，经验不会导致真正的信念，而只会改善观点。多对一问题仍然存在。正如你有许多图像版本一样，也有许多实体猫，每只猫都是独一无二的，但我们用一个术语“猫”来概括它们。

Seeing a person or an object is a step-up Socrates’ ladder of knowledge, causing beliefs rather than imaginings. But sensations create perceptions, mental pictures of objects in the world, and sensation is a bodily process not to be trusted. Think of two people entering a heated room one winter night. Person A has been downstairs stoking a coal furnace where the temperature was 100 degrees. Person B enters from outside where the temperature was 0 degrees. A will find the room pleasantly cool, while B will find it pleasantly warm. But which is it, cool or warm? A and B sense the room differently because of the different physical states of their bodies: Each knows their private experience; neither senses the True temperature of the room. Because experience depends on the state of our (unreliable physical) bodies, experience does not cause true belief, only improved opinion. The many-to-one problem remains, too. Just as there are many image versions of you, there are many physical cats each of which is unique, yet we apply one term, CAT, to all of them.

此时，柏拉图和苏格拉底的许多学生陷入了困境，因为似乎无法弥合可错经验与认识论真理之间的鸿沟。事实上，苏格拉底的大多数追随者认为，真正的知识并非人类所能获得，于是他们成为了怀疑论者（不完全相信任何事物）、愤世嫉俗者（不信任社会），以及后来的实用主义者（真理就是有用的，不追求真理）。

At this point, Plato and many of Socrates’ students faced an impasse because there seemed no way to overcome the gap between fallible experience and epistemological Truth. Indeed, most of Socrates’ followers concluded that genuine knowledge was not attainable by human beings and became skeptics (don’t believe anything 100%), cynics (don’t trust society), or, later, pragmatists (truth is what works, and don’t ask for Truth).

然而，学习毕达哥拉斯几何学却让柏拉图解答了苏格拉底怀疑论的难题：存在一个由普遍性构成的非物质现实，它们可以被非物质的理性灵魂所认知。在高中几何课上，你可能证明了勾股定理，即以直角三角形斜边（c ）为界的正方形面积等于其余三边（ a和b ）的平方和，即 a2 ⁺ b2 ⁼ c2 ^。你并没有依靠观察来证明这个定理。你可能画了一个三角形来帮助你思考，但它在逐步展开证明的过程中并没有发挥任何作用。你也没有使用科学方法。你没有从同学那里收集 30 张三角形图像，仔细地在各个边上竖立正方形，测量它们，计算它们的面积，并检查这些值是否符合方程 a2 ⁺ b2 ⁼ c2 ^，在某个测量标准误差 p < 0.05 的范围内^。2

Learning Pythagorean geometry, however, gave Plato an answer to the Socratic skeptical conundrum: There is a non-physical reality populated by universals, and they are knowable by the non-physical rational soul. In high school geometry you probably proved the Pythagorean Theorem, that the area of a square erected on the hypotenuse (c) of a right-angled triangle is equal to the sum of the squares of the remaining sides (a and b), that is a² + b² = c². You did not rely on observations to produce the proof. You may have drawn a triangle to help you think, but it played no part in the step-by-step unfolding of the proof. Nor did you use the scientific method. You did not collect 30 images of triangles from your classmates, carefully erect squares on all sides, measure them, calculate their areas, and check to see if the values fit the equation a² + b² = c² within some standard error of measurement, p < 0.05.²

你的证明是基于几何学的第一原理或公理的逻辑推导，没有受到任何感觉的污染。简而言之，这是纯粹理性和逻辑灵魂的体现。对柏拉图来说，最重要的是，你提供的是一个真正的证明，它确立了勾股定理并非一个可错的观点，而是一个必然真理，对任何能够理解其逻辑的人来说都是令人信服的。苏格拉底已经表明，知识本身并非信念比观点更需要有理有据。柏拉图现在断言，理据是逻辑证明，而非经验观察。

Your proof was a logical deduction from the first principles, or axioms, of geometry, uncontaminated by sensation. In short, it was a pure act of the rational, logical soul. Most importantly for Plato, what you produced was a real proof that established the Pythagorean Theorem not as a fallible opinion but as a necessary Truth compelling to anyone capable of following its logic. Socrates had shown that to be knowledge rather than opinion a belief had to be justifiable. Plato now asserted that justification was logical proof rather than empirical observation.

然而，多对一的问题依然存在。你画的三角形，你同学画的三角形，或者任何人画的三角形，都是一个固定不变的直角三角形的有缺陷的复制品。柏拉图大胆地宣称，这种众多个体与一个真实模型（或形式）的关系适用于世间万物。

The many-to-one problem, however, remained. Your drawn triangle, your classmates’ drawn triangles, or anybody’s drawn triangles are individual flawed copies of a single fixed and eternal, Right-Angled Triangle. Plato boldly proclaimed that this relationship of many individuals to one True model, or Form, applies to everything in the world.

柏拉图给他的继承者们留下了一个指导框架，用于思考观点与知识之间、人类思想与身体之间以及追求真理之间的关系。

Plato bequeathed to his successors a guiding framework for thinking about the relationships between opinion and knowledge and between the human mind and body and the pursuit of Truth.

认知科学的四种经典立场

Four Classical Positions in Cognitive Science

现在，我们可以把焦点从知识问题——什么是知识？——转移到认知科学中关于知识的问题——人类思维是如何获取知识的？我们将从大多数人认为是常识的认知——即知者与被知者之间的关系——开始。

We can now shift our focus from questions about knowledge—what is knowledge?—to the cognitive science question about knowledge—how does the human mind acquire knowledge? We will begin with an account of cognition—the relation between knower and known—that most people hold as a matter of common sense.

经验主义 1.0（亚里士多德）：感觉提示：真理是通过人类思维的认知操作从感觉中感知提取出来的

Empiricism 1.0 (Aristotle): Sensation Hints: Truth Is Perceptually Extracted from Sensation by the Cognitive Operations of the Human Mind

亚里士多德（公元前384-322年）虽然师从柏拉图，但他本质上是一位科学家。他的座右铭是让事实指导我们的理论，解决我们的争议。他没有老师那种宗教和神秘主义的一面。因此，在拉斐尔的名画《雅典学院》（约公元1510年）中，柏拉图指向上方的形式，在那里可以找到真理，而亚里士多德则指向下方，指向我们世界的实践知识。

Although Aristotle (384–322 BCE) studied with Plato, he was fundamentally a scientist. His motto was to let facts guide our theories and settle our disputes. He did not have his teacher’s religious cum mystical side. Thus, in Raphael’s famous painting The School of Athens (c. 1510 CE), Plato points upward to the Forms where Truth is to be found, while Aristotle points down to practical knowledge of our world.

亚里士多德相信感官，但他意识到朴素实在论过于相信经验。因此，他提出了一种对人类经验的心理学解释，以证明他的哲学主张：正确运用感知可以证明真实的信念（图1.2）。

Aristotle trusted the senses but realized that naïve realism put too much faith in experience. Therefore, he proposed a psychological account of human experience to justify his philosophical claim that rightly used, perception could justify true beliefs (Figure 1.2).

从认识论的角度来看，他的关键举措是将感觉与知觉区分开来。德谟克利特的朴素实在论断言，当一个物体的幻象进入我们的血液时，我们会在瞬间在精神上把握住这个物体。然而，亚里士多德认为经验是一个漫长的建构和解释过程。它始于特殊的，或者更好的说法是专门的视觉、听觉、触觉、味觉和嗅觉，可以记录感知对象的特定方面。

From an epistemological perspective, his key move is separating sensation from perception. Democritus’ naïve realism asserted that when the eidolon of an object enters our bloodstream, we mentally grasp that object in a single moment. Aristotle, however, sees experience as a drawn-out constructive and interpretive process. It begins with the special, or, better, specialized, senses of seeing, hearing, touching, tasting, and smelling, which register particular aspects of a perceived object.

该物体尚未被分类；这项工作留给常识，即各种特殊感官的输出组合成一幅完整物体的图像。在提出常识时，亚里士多德正在处理认知神经科学中所谓的绑定问题。视觉始于视网膜中的光激活细胞，它们的神经输出被处理到大脑后部的视觉皮层。听觉始于耳蜗中细胞的刺激，它们的输出被处理到听觉皮层，其他感官也是如此。尚未解决的问题是大脑如何将这些独立的区域绑定在一起形成意识。

The object is not yet categorized; that job is left to common sense, where the outputs of the special senses combine into a picture of the complete object. In positing common sense, Aristotle is dealing with what’s called in cognitive neuroscience the binding problem. Vision begins with light activating cells in the retina whose neural outputs are processed to the visual cortex in the back of the brain. Hearing begins with stimulation of the cells in the cochlea, whose output is processed to the auditory cortex, and so on for the other senses. The still unsolved problem is how the brain binds these separate areas together into consciousness.

因此，感觉——对感官信息片段的记录——与知觉（将一组感官识别为猫或狗）是分离的且先于知觉。这种心理学上的区别使亚里士多德能够为经验作为通往知识的途径进行辩护。这里有一个基于亚里士多德的例子。我的女儿贝丝在弗吉尼亚联邦大学主修艺术，我们很自然地会在校园里见面。然而，有另一个艺术系的学生长得非常像贝丝，有时我差点挥手叫她的名字，却发现我看到的人不是贝丝，而是她的分身。乍一看，这个例子似乎支持对经验准确性的怀疑，但亚里士多德指出，我的特殊感官都是正确的：我看到了一个特定身高、特定头发颜色等等的年轻女子。错误在于我做出了认知判断，认为常识拼凑在一起的感官片段构成了贝丝。

Sensation—the registering of pieces of sensory information—is thus separate from and prior to perception, the recognition of a collection of sensations as a cat or a dog. This psychological distinction allows Aristotle to mount a defense of experience as a path to knowledge. Here’s an example based on one given by Aristotle. My daughter Beth was an art major at Virginia Commonwealth University, and we would naturally see each other around campus. However, there was another art student who looked very much like Beth, and sometimes I would nearly wave and call Beth’s name only to realize that the person I saw was not Beth but her doppelganger. At first glance, this example seems to support skepticism about the accuracy of experience, but Aristotle pointed out that my special senses were all correct: I saw a young woman of a particular height, a particular hair color, and so on. The error was the cognitive judgment I made that the sensory pieces assembled together by common sense constituted Beth.

与柏拉图相反，原始感觉是准确可靠的；错误知觉并非源于感觉，而是思维根据组合起来的知觉进行推理时产生的错误。常识在我们心中构建了由特殊感官的输出组合而成的经验对象的表征，我们用这些表征进行思考，并思考这些表征。柏拉图拒绝经验太过极端。我们需要警惕的是大脑根据输入得出的结论，而不是输入本身。

Contra Plato, raw sensation is accurate and reliable; misperception occurs not in sensation, but when thinking makes inferential errors from the assembled percept. Common sense creates in us representations of experienced objects assembled from the outputs of the special senses, and we think with and about these representations. Plato’s rejection of experience was too extreme. We need to beware the conclusions our minds reach given its inputs, not the inputs themselves.

感知到的物体的表征会被传递到想象力，想象力指的是心灵即使在没有物体的情况下也能将感知保持在意识中的能力。然后，图像可以存储在记忆中，以供日后检索。常识、想象力和记忆构成了内在感官，因为它们都与感官体验的各个方面有关，无论是第一手的（常识），还是短期的（想象）或长期的（记忆）。动物和人类的心灵都具有这些能力。动物，即使是我们最亲近的亲戚——类人猿，也会在发现和记忆世界时了解和应对它，但它们不会构建普遍的概念（Tomasello，2019）。猫会追逐老鼠、田鼠和鼩鼱，还会追逐弹珠和激光点，但它们不会创造这些概念。

Representations of perceived objects are passed to the mental faculty of imagination, which refers to the ability of the mind to hold a percept in consciousness even in the absence of the object. The image can then be stored in memory, becoming available for later retrieval. Common sense, imagination, and memory comprised the interior senses because they all dealt with aspects of sensory experience, either at first hand (common sense) or in short-term (imagination) or long-term (memory) memory. They are shared by animal and human minds. Animals, even our closest relatives, the great apes, know and cope with the world as they find it and remember it, but they do not construct universal concepts (Tomasello, 2019). Cats chase mice, voles, and shrews, plus marbles and laser dots, but they don’t create the concepts.

亚里士多德与柏拉图都认为，从有限的个体经验中形成普遍概念的能力是人类心灵的独特力量。他将心灵分为两种能力：主动心灵和被动心灵。粗略地说，主动心灵包含一系列过程，这些过程会记录不同个体之间的异同，构建每个个体的程式化图像，然后将其存储在被动心灵中。普遍性知识必须通过学习获得。

Aristotle shared with Plato the idea that the ability to form universal concepts from the limited experience of individuals is the unique power of the human mind. He divided mind into two faculties, active mind and passive. Roughly speaking, active mind comprises a set of processes that note similarities and differences between various individual cats, dogs, and so on, building up stylized pictures of each that are then stored in passive mind. Knowledge of universals must be learned.

亚里士多德的认知心理学采用了现代认知科学的一个关键策略，即解释分解。以记忆为例。我们（和动物）会记住事物，但亚里士多德并没有提出一个单一的记忆系统，而是提出了两个以不同方式组织起来的记忆系统：动物记忆和被动思维。它们类似于一种重要的现代记忆类型学（Tulving，1972），介于情景（或自传体）记忆和语义记忆（或知识）之间。动物记忆储存了由常识和想象力整合的先前经验。当被要求回忆一段个人记忆时，比如“你的第一只宠物是什么？”，你会回溯多年，找到一张比如你的宠物狗乔的图像。情景记忆的组织原则是时间，找到正确时代的图像来回忆。另一方面，如果我问你什么是狗，时间并不重要。相反，被动思维的组织原则是意义——能够将乔归类为狗和宠物这两个普遍概念的成员。重要的是，我们发现，与我们的直觉相反，我们并没有一个包罗万象的记忆系统，而是至少有两个。

Aristotle’s cognitive psychology adopts a key strategy of modern cognitive science, explanatory decomposition. Consider memory. We (and animals) remember things, but instead of positing a single MEMORY system, Aristotle posits two, which are organized in different ways: Animal memory and passive mind. They resemble an important modern typology (Tulving, 1972) of memory, between episodic (or autobiographical) memory and semantic memory (or knowledge). Animal memory stores previous experiences integrated by the faculties of common sense and imagination. Asked to recall a personal memory, such as “What was your first pet?”, you search back through the years to find an image of, say, your pet dog, Joe. The organizing principle of episodic memory is time, finding the right era’s images to recall. On the other hand, if I ask you what a dog is, time is unimportant. Instead, the organizing principle of passive mind is meaning—being able to classify Joe as a member of the universal concepts DOG and PET. Importantly, we see that, contrary to our intuition, we don’t have one all-encompassing memory system, but at least two.

一些必要的形而上学

Some Necessary Metaphysics

我曾试图回避形而上学，但形而上学的思想引发了人们对人类在自然界中地位的普遍反思，并使认知心理学不仅成为一门可能的科学，而且成为一门重要的科学。第一个变化，关于共相的本质（多对一问题），削弱了认识论的权威，反过来提升了心理学的权威。第二个变化，认识到许多人类感觉是精神创造而非物理事实，创造了一个新的探索世界——意识的内在世界。第三个变化，机器思维的兴起，创造了现代科学和人工智能。

I have tried to avoid metaphysics, but metaphysical ideas set in motion a general rethink of the place of humans in nature, and they made cognitive psychology not just a possible science, but also an important one. The first change, about the nature of universals (the many-to-one problem) drained epistemology of its authority, elevating, in turn, the authority of psychology. The second, the realization that many human sensations were mental creations rather than physical facts, created a new world for exploration—the inner world of consciousness. The third, the rise of machine-thinking, created modern science and artificial intelligence.

世界图景的机械化

The Mechanization of the World Picture

两种不同却又兼容的隐喻主导着古代关于宇宙的思想：宇宙如书，宇宙如生命。柏拉图主义鼓励使用书籍隐喻。感知世界是形式的一种表达，正如我们阅读散文是为了领悟作者的内在含义，我们也可以阅读物理对象的“散文”，以辨别其含义，将其作为通往隐藏形式的线索。斯多葛学派认为，宇宙是一个随着时间而发展的生命体。它的发展是预先确定的，而我们最幸福的方式就是接受发生在我们身上的一切，将其视为神圣进程的一部分。

Two different, compatible, metaphors dominated ancient thinking about the universe: Universe as book and universe as living being. Platonism encouraged the book metaphor. The perceived world was an expression of the Forms, and just as we read prose to discern an author’s inward meaning, so we can read the “prose” of physical objects to discern their meaning as clues to the hidden Forms. Stoicism taught that the universe was a living being developing over time. Its development was pre-determined, and we can best be happy by accepting what happens to us as part of a divine process.

科学革命的核心是用机器的隐喻取代了宇宙的书籍/存在隐喻。自希腊和罗马时代以来，工匠们就一直在制造机器，但这些机器主要用于娱乐或宗教目的。很大程度上是因为奴隶可以从事繁重的体力劳动，机器并没有被用来提高经济生产力。在文艺复兴时期和近代早期——科学革命时期——娱乐性机器制造蓬勃发展，从贵族宅邸和花园中的自动机到四处游荡的机械棋手（Riskin，2003；2016）。

Central to the Scientific Revolution was replacing the book/being metaphor for the universe with the machine metaphor. Artisans had been building machines since Greek and Roman times, but they were built for entertainment or religious purposes. Largely because of the availability of slaves to do heavy work, machines were not deployed to increase economic productivity. In the Renaissance and early modern period—the time of the Scientific Revolution—there was an explosion of entertaining machine-building, from automata populating aristocratic mansions and gardens to a traveling mechanical chess player (Riskin, 2003; 2016).

对机器的熟悉启发了我们思考宇宙、动物身体，甚至行为和思维的新视角。科学将宇宙描述为一个巨大的钟表，由上帝创造并启动，但其行为却由其自身的运作自主支配。

Familiarity with machines suggested new ways to think about the universe, animal bodies, and even behavior and mind. Science described the universe as an immense clockwork created and set in motion by God, but whose behavior was governed autonomously by its own workings.

机器隐喻影响深远。在一个宗教战争四分五裂的时代，科学家们可以研究和构建物质世界理论，而无需触犯神学的痛处。无论创造宇宙的上帝本质如何，他都将其创造得完美无缺，并使其按照固定不变的永恒法则运转，这些法则可以被科学家和其他人类观察、记录并用于人类的目的。机器隐喻揭开了书籍隐喻的神秘面纱，因为宇宙时钟受数学定律（例如牛顿运动定律和万有引力定律）的支配，因此所有未来事件，例如哈雷彗星的回归，都可以被精确预测。牛顿定律的数学意义不容小觑。如果牛顿只是简单地说苹果会掉落，没有人会对此感到惊奇。但能够预测苹果掉落的确切时间，则表明了预测和控制事件的可能性。钟表宇宙最大的吸引力在于，了解它的数学弹簧和齿轮意味着人类不再需要在自然面前感到被动，而是能够运用来之不易的科学知识来控制宇宙这台机器，服务于人类的目标。正如科学革命的思想家弗朗西斯·培根（1561-1626）所言：“知识就是力量。”

The machine metaphor was influential. In an age fractured by religious warfare, scientists could study and theorize about the physical world without stepping on sore theological toes. Whatever the nature of the God who had fashioned the universe, He had made it perfect and set it in motion governed by fixed, eternal rules that could be observed, written down, and used by scientists and other human beings for human ends. The machine metaphor removed the mystery of the book metaphor, as the universe-clock was governed by mathematical laws such as Newton’s Laws of Motion and Gravity so that every future event, such as the return of Halley’s comet, could be precisely predicted. The mathematics of Newton’s laws should not be underestimated. If Newton had simply said that apples fall, no one would have been impressed. But being able to predict the exact duration of an apple’s fall suggested the possibility of anticipating and controlling events. The greatest appeal of the clockwork universe was that to know its mathematical springs and gears meant that human beings need no longer feel passive in the face of nature but could use hard-won scientific knowledge to control the universe-machine to serve human ends. As Francis Bacon (1561–1626), the ideologist of the Scientific Revolution, wrote, “Knowledge is power.”

由于当时许多流行的机械自动机模仿的是生物，动物，甚至人类，可能是机器的想法逐渐受到关注。医学研究进一步证实了这一观点，解剖揭示了人体是机器，即由相互连接的物理部件组成的更大整体。然而，动物身体与人造机器的类比并不清楚。动物仅仅是钟表，行为像钟表一样可预测的机器吗？或者，动物只是像钟表一样，由物理部件组成，却拥有某种人类机器所缺乏的自主生命力，能够自我调节。机器能够以时钟无法做到的方式适应不断变化的环境吗？如果真是这样，工程师们能否制造出像动物一样，拥有学习能力和自主行为能力的机器？(Riskin, 2016)

Because many of the popular mechanical automata of the day imitated living things, the idea that animals, and perhaps humans, might be machines gained traction. Research in medicine reinforced the notion when dissection revealed that bodies were machines, that is, collections of interconnected physical parts comprising a larger whole. However, the analogy of animal body to human-made machine was ambiguous. Were animals just clockwork, machines whose behavior was as predictable as a clock? Or were animals simply like clockwork, made of physical parts yet possessing something, an autonomous life-force, that human machines lacked, able to adjust themselves to changing environments in a way clocks can’t? If so, could engineers build machines that were like animals, capable of learning and autonomous behavior (Riskin, 2016)?

机械身体的含义通过最著名的自动机之一——机械鸭子（Riskin，2003b）（图 1.3）戏剧性地体现出来，它是由自动机界的元老雅克·德·沃康松（Jacques de Vaucanson，1709–1782）制造的。

The implications of machine bodies were made dramatically manifest by one of the most famous of all the automata, the mechanical duck (Riskin, 2003b), Figure 1.3, built by the doyenne of automatists, Jacques de Vaucanson (1709–1782).

鸭子使用弹簧、齿轮、管子和滑轮来行走和嘎嘎叫。然而，沃康松增加了一个新技巧：他的鸭子看起来像鸭子，走路像鸭子，叫得也像鸭子，但它排便也像鸭子（具体方式尚不清楚）。将肌肉运动视为机械运动很简单，但将消化和废物清除视为机械运动则更具挑战性，因为种子和粪便之间存在质的变化。然而，鸭子（似乎）做到了，而且它是人造机器，而不是有机生长的动物。如果动物的消化器官（真实的和机械的）可以产生粪便，那么也许动物和人类的大脑可以“像肝脏分泌胆汁一样分泌思想”，正如法国医生皮埃尔·卡巴尼斯（1757-1808）后来声称的那样。人工消化意味着人工智能。

The duck used springs, gears, tubes, and pulleys to make it walk and quack. However, Vaucanson added a new trick: His duck looked like a duck, walked like a duck, and quacked like a duck, but it also defecated like a duck (by means still unclear). Thinking of muscular movement as mechanical is straightforward, but thinking of digestion and waste removal as mechanical is more challenging because there’s a qualitative change between seeds and feces. Yet the duck (appeared) to do it, and it was a human-built machine, not an organically grown animal. If animal digestive organs, real and mechanical, could produce feces, perhaps animal and human brains can “secrete thought as the liver secretes bile,” as French physician Pierre Cabanis (1757–1808) claimed later. Artificial digestion implied artificial intelligence.

哪些感觉是真实的？意识的创造

Which Sensations Are Real? The Creation of Consciousness

正如古典思想家们认为人类概念与外部形式相对应一样，他们也认为人类的每一种感觉都与感知对象的真实特征相对应。德谟克利特的朴素实在论者认为，幻象是他们留下的物体的复制品，因此如果我们体验到的是红苹果，那是因为苹果幻影是红色的。同样，亚里士多德认为，眼液会呈现出所见物体的颜色。

Just as they assumed that human concepts correspond to external Forms, classical thinkers supposed that every human sensation corresponds to a real feature of a sensed object. Democritean naïve realists thought that eidola were copies of the objects they left, so that if we experience a red apple, it was because the apple eidolon was red. Similarly, Aristotle thought that the fluid of the eye takes on the colors of the objects it sees.

然而，将世界视为一台类似时钟的机器，却让科学家们对这一合理的假设产生了怀疑。他们想要一个物理世界的机械解释，在这个物理世界中，物体凭借其物理属性（例如质量、动量和数量）发生因果相互作用，这些属性是可量化的，并且能够被代入方程式。但并非所有感官属性都具备这些特征。伽利略在其论文《试金者》（1623）中写道：

However, viewing the world as a clock-like machine caused scientists to doubt this plausible assumption. They wanted a mechanical account of the physical world in which objects causally interact by virtue of their physical properties, such as mass, momentum, and number, properties that were quantifiable and capable of being entered into equations. But not all sense properties have these characteristics. Galileo wrote in his essay, The Assayer (1623):

每当我构想任何物质或实体时，我……都会将其视为有界的，具有各种形状；有大有小，有动有静……我无法用任何想象力将某种物质与这些状态区分开来。但它必须是白色或红色、苦或甜、嘈杂或寂静、散发着甜臭或恶臭，我的大脑并不会被迫将其作为必要的伴随物引入……因此，我认为味道、气味、颜色等等……只存在于意识中，如果生物被移除，所有这些特质都会被抹去并消灭。

Smith 于 1990 年引用，第 739 页

Whenever I conceive any material or corporeal substance I… think of it as bounded, and as having this or that shape; as being large or small [and] as being in motion or at rest… From these conditions I cannot separate such a substance by any stretch of my imagination. But that it must be white or red, bitter or sweet, noisy or silent, and of sweet or foul odor, my mind does not feel compelled to bring in as necessary accompaniments…. Hence, I think that tastes, odors, colors, and so on… reside only in the consciousness [so that] if the living creature were removed all these qualities would be wiped away and annihilated.

quoted by Smith, 1990, p. 739

伽利略在这段文字中做出了非凡的举动。首先，他的目标是形而上学的，旨在提出这样一种观点：物理世界是一台拥有少量物理属性的机器，这些属性规律地相互作用，使钟表宇宙自行运转，无需上帝或巫术的干预。然而，他将感知到的物理属性（现在称为“主要感知属性”）与心灵添加的属性（次要感知属性）区分开来的论点，则带有心理学色彩。

Galileo makes remarkable moves in this passage. First, his goal is metaphysical, to present the idea that the physical world is a machine possessing a small number of physical properties that interact lawfully to make the clockwork universe operate by itself without intervention by God or sorcery. However, his argument separating sensed physical properties (known now as the primary sense properties) from those that the mind adds (the secondary sense properties) is psychological.

此外，他的心理学论证建立在内省之上。从朴素实在论者到斯多葛学派，古典思想家们都认为心灵的存在是理所当然的，但他们从未探索过自己的心灵，也没有基于自身经验提出重要论断（例如，宇宙是一台机器）的论证。

Moreover, his psychological argument rests on introspection. From the naïve realists to the Stoics, the Classic thinkers took the existence of mind for granted, but they never explored their own minds, nor did they propose arguments for important claims (e.g., that the universe is a machine) based on their own private experience.

最后，在《试金者》中，伽利略引入了一个在古代心理学著作（罗蒂，1979）中找不到的词：意识。研究意识将成为认知科学的起点：如果像颜色这样的次要感官属性并非像实在论者和经验论者所认为的那样，简单地从自然中习得，那么我们为什么会看到颜色呢？伽利略关于哪些感官属性是主要、哪些是次要的，是否正确？

Finally, in The Assayer, Galileo introduces a word that we don’t find in ancient psychological writing (Rorty, 1979): consciousness. Investigating consciousness will be the starting point of cognitive science: If a secondary sense property such as color is not simply picked up from nature, as realists and empiricists supposed, why do we see colors? Is Galileo right about which sense properties are primary and secondary?

正如科学革命期间的许多其他事物一样，一种新机器——暗箱（图 1.4）——提供了一个隐喻，这次是指意识。

As in so much else during the Scientific Revolution, a new machine, the camera obscura (Figure 1.4), offered a metaphor, this time for consciousness.

它已经为眼睛提供了一个模型，它与眼睛非常相似，其聚焦透镜将物体的图像投射到盒子的背面。艺术运用为隐喻诠释增添了新的层面。为了捕捉线性透视，艺术家使用暗箱在画布上投射出他们所描绘物体的二维图像。想要画画。他们利用投影进行炭笔素描，捕捉深度感知的线索，然后画下他们的画作（Andersen，2013）。作为意识的隐喻，画家成为内在自我——伽利略的自我——而画布则是意识，是自我与世界之间的屏障。这道屏障是如何产生的，成为了认知科学的研究主题。由于智能，即自我，如何运作，成为一个令人望而生畏的谜团，因为它并没有被投射到屏幕上。自我变成了一个黑洞，认知科学围绕着它运作，却无法解释。

It had already provided a model for the eye, which it closely resembles with its focusing lens projecting an object’s image on the back of the box. Artistic use added a new layer for metaphorical interpretation. In order to capture linear perspective, artists used cameras obscura to cast on canvas a two-dimensional image of the objects they wanted to paint. They used the projection to make a charcoal sketch, capturing cues to depth perception, and then painted their picture (Andersen, 2013). As a metaphor for consciousness, the painter becomes the inner self—Galileo’s self—and the canvas is consciousness, a screen between self and world. How the screen is created became the subject matter of cognitive science. How intelligence, the self, worked, became a forbidding mystery because it was not projected on the screen. The self became a black hole around which cognitive science operated but could not explain.

科学家们似乎发现了一个精神的新世界——意识。认知心理学探索意识的蓝图是由勒内·笛卡尔（1596-1650）绘制的。他将科学革命的各种倾向和隐喻整合、提炼、强化和整理，形成了一种关于心智、身体、意识和智力的思考方式，这种思考方式在认知科学领域主导了两个多世纪。

Scientists seemed to have discovered a mental New World, consciousness. The chart for its exploration by cognitive psychology was set by René Descartes (1596–1650), who pulled together, refined, strengthened, and ordered the various tendencies and metaphors of the Scientific Revolution into a way of thinking about mind, body, consciousness, and intelligence that dominated cognitive science for more than two centuries

笛卡尔范式：机械身体中的非物质灵魂

The Cartesian Paradigm: The Immaterial Soul in the Mechanical Body

笛卡尔生活于欧洲历史上一个变革的世纪。从政治到地球的形状，一切都悬而未决。科学与神学、政治、权力和战争相互交织。任何深刻思想家的思想都需要仔细的阐释，但笛卡尔的需要出于特殊的原因。我们知道，在得知伽利略被宗教裁判所判罪后，他压制了自己的物理学著作《世界》 ，并停止了其姊妹篇《人》的创作。他放弃这些思想是因为它们是异端邪说吗？他是否暂时压制它们，直到他能够将它们与他的天主教信仰相协调？他是否……保留它们，却又用各种逃避的借口来掩饰？我们无法确定。在阐述我所谓的笛卡尔范式时，我将强调笛卡尔对认知科学影响最大的诠释，即使这些诠释可能并非笛卡尔的本意（Leahey，2018；Riskin，2016）。

Descartes lived in a transformative century in European history. Everything from politics to the shape of the globe was up for grabs. Science rubbed up against theology, politics, power, and war. Any profound thinker’s ideas need careful exegesis, but Descartes’ need it for special reasons. We know he suppressed his work on physics, The World, and stopped working on its companion, The Human, after learning of Galileo’s condemnation by the Inquisition. Did he give up his ideas because they were heretical? Did he temporarily silence them until he could reconcile them with his Catholic faith? Did he keep them, yet hide them in a bodyguard of evasions? We cannot be sure. In setting out what I call the Cartesian Paradigm, I will emphasize the interpretations of Descartes that had the greatest influence on cognitive science, even if they may not have been exactly what Descartes meant (Leahey, 2018; Riskin, 2016).

动物的身体是一台机器

他那个时代那些精巧的自动机为笛卡尔关于生物体运作的理论提供了模型。他将自动机的齿轮、轮子和弹簧具体化为生理学概念，其中最重要的就是感觉运动反射。

The clever automata of his time provided models for Descartes’ theories about how living bodies operate. He literally fleshed out automatists’ gears, wheels, and springs into physiological concepts, the most important of which was the sensory-motor reflex.

图 1.5中的人的脚 (B) 离火 (A) 太近了，火会拉动笛卡尔所说的“绳子”(C)，该绳子连接到大脑的松果体 (F)。松果体感受到拉力，反射性地向后拉，将脚从火中拽出。笛卡尔的反射并不依赖于思维——它是一种与生俱来的、预先设定的回避反应。目前尚不清楚这种反射是否依赖于疼痛感。在笛卡尔时代，批评家（以及一些笛卡尔的追随者，尽管不是笛卡尔[Cottingham, 1978]）断言它并不依赖于疼痛感，并得出结论：动物不会感到疼痛，只是会做出类似疼痛的行为。在现代认知科学方面，关于哲学僵尸（没有感觉但行为像正常人一样的生物⁴ ）的猜测很多，或者感觉体验对于适应性行为是否如此重要以至于计算机永远不可能变得智能。

The person in the drawing in Figure 1.5 has got his foot (B) too close to a fire (A), which pulls what Descartes calls a “string” (C), connecting to the brain’s pineal gland (F). Feeling the pull, the pineal gland reflexively pulls back, yanking the foot from the fire. Descartes’ reflex does not depend on thought—it’s an innate, pre-wired, avoidance response. It’s not clear whether the reflex depends upon sensations of pain. In Descartes’ time, critics (and some of Descartes’ followers, though not Descartes [Cottingham, 1978]) asserted that it did not, concluding that animals don’t feel pain, but only act as if they do. In modern cognitive science, speculations abound about philosophical zombies—creatures with no sensations but who act like normal people⁴—or whether sensory experience is so important to adaptive behavior that computers can never be intelligent.

非物质灵魂

笛卡尔抛弃了过时的古代思想，在自己内心寻找哲学和科学的新基础。

Discarding ancient thought as obsolete, Descartes looked inside himself for new foundations for philosophy and science.

他始于苏格拉底的怀疑论，系统地怀疑每一种信仰、观点、知识，无论它有多么珍贵。我可能会问，上帝存在吗？我凭着信念而不是证据来接受它，所以我可以怀疑上帝的存在。我真的坐在舒适的火炉旁吗？不，因为我可能在做梦。笛卡尔继续怀疑许多事情，但发现在他所有的怀疑中，在怀疑的行为中，总有一些东西。他不能怀疑自己怀疑，因为那是自相矛盾的。怀疑是一种思想形式，笛卡尔的结论是，有一件事是绝对的、不容置疑的、必然的真理：我思故我在（Cogito, ergo sum）。

He began with Socratic skepticism, systematically doubting every belief, opinion, piece of knowledge, no matter how precious it might be. Does God exist? I might ask. I take it on faith, not evidence, so I can doubt that God exists. Am I really sitting here next to a comfortable fire? No, because I might be dreaming. Descartes continued to doubt many things, but found that something abides amongst all his doubting, the act of doubting. He cannot doubt that he doubts, for that is self-contradictory. Doubt is a form of thought, and Descartes concluded that one thing is absolutely, indubitably, and necessarily True: I think, therefore I am (Cogito, ergo sum).

他也将自己的真理表述为“我是一个会思考的东西”。然而，这远远超出了他的内省所确立的范围（James, 1904）。他已经表明，意识思维过程贯穿于各种经验之中，但他并未确定进行思考的东西——灵魂或思想——的存在，尽管他认为自己已经确定了。笛卡尔玩弄了智力上的花招，甚至欺骗了自己，他制造了认知科学的最大挑战和最邪恶的陷阱，即意识和创造性智慧与机械身体之间看似不可逾越的鸿沟。

He also put his truth as I am a thing that thinks. However, this goes well past what his introspection established (James, 1904). He’s shown that processes of conscious thought persist through all kinds of experience, but he has not established the existence of a thing that does the thinking—a soul or mind—although he believes he has. In what amounted to an act of intellectual sleight of hand that fooled even himself, Descartes created cognitive science’s greatest challenge and its most nefarious trap, the seemingly insurmountable divide between consciousness and creative intelligence, on one hand, and the mechanical body, on the other.

比较认知：身体与灵魂

宗教与科学的共同作用，使得比较动物与人类的心理能力比亚里士多德之前认为的更加困难，也更加重要。亚里士多德认为所有生物都拥有灵魂，因此也拥有心理功能，因此他能够接受从动物到人类认知能力的连续性。然而，笛卡尔的宗教认为动物没有灵魂，而他的科学则认为动物是机器，对外界刺激做出反射性而非深思熟虑的反应。他需要对动物和人类心理做出新的解读。

Religion and science conspired to make comparing the psychological abilities of animals with those of humans more difficult and more important than it had been for Aristotle. Thinking that all living beings possess souls, and thus psychological functions, Aristotle could accept a continuum of cognitive capacities from animals to people. However, Descartes’ religion taught that animals don’t have souls and his science taught that they are machines responding reflexively, not thoughtfully, to external stimuli. He required a new picture of animal and human psychology.

笛卡尔说，动物和人类在三个方面存在越来越根本的区别。

Descartes said that animals and humans differed in three increasingly more fundamental ways.

• 意识与自我意识。笛卡尔并没有否认动物会感到疼痛，也没有声称它们根本没有感觉。他说，动物拥有简单的意识了解周围的世界，并像我们“心不在焉时”一样自动对刺激做出反应。想象一下，你一边开车走在一条熟悉的路上，一边和朋友聊天。你的思绪沉浸在谈话中，但你驾驶得正确，红灯时停车，在大街上右转，等等。你的动物身体能够感知周围的环境并做出反射性的反应。但这种简单的意识无法像笛卡尔在冥想中那样自我反思。做到这一点需要怀疑理性的全部关注。动物仅限于简单的意识；人类由于是动物而拥有简单的意识，但由于拥有灵魂或思想而拥有反思性的自我意识。
• Awareness vs. Self-Awareness. Descartes did not deny that animals felt pain or claim that they had no sensations at all. He said that animals possess simple awareness of the world around them, and respond automatically to stimuli as we do, “when our minds are elsewhere.” Imagine you are driving a familiar route while talking to a friend. Your mind is engaged with the conversation, yet you drive correctly, stopping for red lights, turning right on Main Street, etc. Your animal body is aware of its environment and responds reflexively to it. But this kind of simple awareness cannot reflect on itself, as Descartes did in his meditations. Doing that required the full attention of skeptical reason. Animals are limited to simple awareness; humans have simple awareness by virtue of being animals but have reflective self-awareness by virtue of having a soul, or mind.
• 条件反射 vs. 灵活适应。现在想象一下，你收到一个驾驶应用程序的紧急通知，告诉你你通常进城的那座桥因事故而关闭。现在你必须——对笛卡尔来说，关键的洞见是你可以——停下来思考。你把思绪从谈话和驾驶中抽离出来，让你的朋友安静下来，然后把车停在路边，评估可用的路线选择。动物无法做到这一点，因为它们被困在一系列自动反射中。人类的大脑可以提前思考；动物的机器却不能。
• Reflex vs. Flexible Adaptation. Now imagine that you get an urgent notice from a driving app, telling you that the bridge you usually take into town is closed because of an accident. Now you must—and for Descartes the key insight is that you can—stop and think. You disengage your thought from talking and driving, shushing your friend and pulling over to assess the routing options available. An animal cannot do this, trapped as it is in a sequence of automatic reflexes. The human mind can think ahead; the animal machine cannot.
• 语言。人类拥有而动物所不具备的关键心理能力是语言。笛卡尔对真理的怀疑论探索，被构建成一系列以“我思故我在”结尾的命题。你与乘客之间那场引人入胜的对话，正是通过语言进行的。当你思考如何在桥上发生事故的情况下到达城镇时，你用脑海中的词语和句子来评估你的选择。语言还能赋予你诸如“猫”、“橡树”或“桥”等普遍概念的名称和描述，而这些概念是动物所不具备的。
• Language. The key psychological ability humans have that animals don’t is language. Descartes’s skeptical search for truth was set out as a series of propositions ending in cogito ergo sum. Your mind-engaging conversation with your passenger was carried out in language. When you thought about how to get to town despite the accident on the bridge, you assessed your options in mental words and sentences. Language also gives you the names and descriptions of universal concepts such as CAT, OAK, or BRIDGE, which animals do not have.

笛卡尔二元论的挑战

笛卡尔提出了我们所见过的最激进的二元论，但其中的一些困难很快就显现出来。

Descartes proposed the most radical dualism we’ve seen, and some difficulties with it became quickly apparent.

首先，笛卡尔部分承认了他人心灵的问题。我知道我是一个会思考的个体，你也知道你是一个会思考的个体，这通过笛卡尔用来证明“我思故我在”的内省论证得以实现。然而，我无法触及你的意识，你也无法触及我的意识。然而，我们有时会将思想归因于他人，有时甚至动物。我们何时有理由这样做呢？笛卡尔认为，语言的使用是思考的产物，因此也是灵魂存在的产物。

One, the problem of other minds was partly recognized by Descartes. I know I am a thinking thing, and you know you are a thinking thing via the introspective argument Descartes deployed to prove cogito ergo sum. However, I have no access to your consciousness, nor do you have access to mine. Nevertheless, we attribute thought to other people and sometimes to animals. When are we justified in doing so? Descartes argued that use of language is the product of thinking, and therefore of the presence of a soul.

值得注意的是，判断计算机是否具备真正人工智能的标准是图灵测试，该测试要求计算机能够进行对话。5 19世纪进化论的出现也带来了^一个相关的问题。对笛卡尔来说，你要么有灵魂，要么没有灵魂；没有中间地带。但进化论意味着早期和晚期生命形式之间的连续性，使得笛卡尔式的思维进化场景^{6不太可能发生。}

Significantly, the standard test for a computer having genuine artificial intelligence is the Turing Test, which requires that a computer be able to carry on a conversation.⁵ A related problem popped up with the arrival of evolution in the 19th century. For Descartes, either you have a soul or you don’t; there is no middle ground. But evolution implies continuity between earlier and later forms of life, making a Cartesian scenario⁶ for the evolution of mind unlikely.

笛卡尔的一位哲学通信者是波西米亚公主伊丽莎白·斯图亚特（1596-1662）。她问笛卡尔灵魂和肉体是如何相互作用的。答案似乎是肯定的：当我把书掉在脚上时，我会感到疼痛；当我抬起右臂时，它会抬起来。然而，伊丽莎白问道，一个纯粹非物质的灵魂如何能够推动机械躯体的杠杆和齿轮？感觉运动反射的机制又如何能够在非物质的灵魂中引发体验？笛卡尔回答说，松果体是相互作用的中心，身体的神经在这里汇聚，接受灵魂的操控。但伊丽莎白并没有因此而平静下来——灵魂是非物质的，而松果体是机械物质。

One of Descartes’ philosophical correspondents was Elizabeth Stuart, Princess of Bohemia (1596–1662), who asked him how soul and body interacted. It certainly seems they do: When I drop a book on my foot it hurts, and when I choose to lift my right arm it goes up. Yet, Elizabeth asks, how can a purely immaterial soul move the levers and gears of a mechanical body, and how can the mechanics of the sensory-motor reflex cause experiences in the immaterial soul? Descartes replied by saying the pineal gland is the locus of interaction, where the nerves of the body come together to be manipulated by the soul, but Elizabeth was not mollified—the soul is immaterial and the pineal gland is mechanical matter.

最后，伊丽莎白的批评中隐含着笛卡尔二元论最深层的难题，即“侏儒”（homunculus，意为“小人”）问题。在一幅基督教图像中，上帝将亚当的灵魂置于他的身体中。亚当的灵魂被描绘成亚当自身的缩小版，在上帝行神迹之后，灵魂将与亚当的身体共存并指引他的身体。然而，灵魂的运作方式并未得到解释。亚当灵魂体内是否还有一个更小的亚当灵魂？而那个亚当灵魂体内是否还有一个更小的亚当灵魂？笛卡尔认为每个人体内都有灵魂，这种观点是传统的，或许能安慰那些寻求来世的人（Brown，2015），但它无法从科学上解释任何事情，因为灵魂的运作方式超出了科学探究的范围^。7

Finally, implicit in Elizabeth’s criticism is the deepest difficulty of Cartesian dualism, the problem of the homunculus (“little man”). In one Christian image, God is shown placing Adam’s soul into his body. Adam’s soul is depicted as a smaller version of Adam himself, and after God’s miracle the soul will dwell with and direct Adam’s body. However, the operations of the soul are not explained. Does Adam-soul have a smaller Adam-soul inside him? And is there a still smaller Adam-soul inside that one? Descartes positing a soul inside every human is traditional, and it may be comforting to those looking for an afterlife (Brown, 2015), but it cannot scientifically explain anything because the operations of the soul are placed beyond scientific inquiry.⁷

摘要：理念之路与笛卡尔剧场

在笛卡尔看来，柏拉图式的ιδεα s 变成了精神理念。图 1.6出自他的《人》，将他的认知模型描绘成一个巨大的暗箱，或者用认知哲学家丹尼尔·丹尼特 (1991) 的话来说，是笛卡尔剧场。

With Descartes, Platonic ιδεα s became mental ideas. Figure 1.6, from his L’Homme, depicts his model of cognition as a grand camera obscura, or, in the phrase of cognitive philosopher Daniel Dennett (1991), the Cartesian Theater.

我们看到眼睛看着一支箭。箭的每个部分，箭头 (A)、箭杆 (B) 和箭羽 (C)，都通过点 1、3 和 5 处的光线记录在视网膜上。神经 2、4 和 6 将机械刺激传递到大脑，将 A、B 和 C 的图像投射为概念 a、b 和 c，投射到松果体的光滑表面上，在那里它们被灵魂 H 理解。我们并不是直接体验世界，而是像电影院的观众一样，通过表征，这是古典认知科学的定义概念之一。此外，正如机器由较小的部件组装成较大的整体一样，简单的想法也可以组合成更大的整体。箭是一台简单的机器，由三部分组成：箭头 (A)、箭杆 (B) 和箭羽 (C)。同样，我们对箭的复杂概念也是由箭头 (a)、箭杆 (b) 和箭羽 (c) 等简单概念组装而成的。伽利略的两个世界由此展现。一个是由简单部分构成的物体组成的物理世界，另一个则反映在由简单观念构成的复杂观念所构成的精神世界中。因此，认知科学的研究方向变成了揭示物理世界如何映射到精神世界，发现支配观念的精神规则，并揭示当物理世界仅具有主要机械属性时，经验的次要属性是如何产生的。

We see eyes looking at an arrow. Each part of the physical arrow, arrowhead (A), shaft (B), and fletching (C), is registered on the retina by light rays at points 1, 3, and 5. Nerves 2, 4, and 6 carry mechanical stimulation to the brain, casting images of A, B, and C, as ideas a, b, and c, on the smooth surface of the pineal gland, where they are comprehended by the soul, H. We do not experience the world directly but as audiences in a movie theater, by representations, one of the defining concepts of classical cognitive science. Moreover, as machines are made of smaller pieces assembled into a larger whole, simple ideas can be combined into larger wholes. An arrow is a simple machine, made up of three parts: Arrowhead (A), shaft (B), and fletching (C). Similarly, our complex idea of an arrow is assembled out of simpler ideas, of the arrowhead (a), of the shaft (b), and of the fletching (c). Galileo’s two worlds are on display. There is a physical world made up of objects constituted of simpler parts, and it is reflected in a mental world made up of complex ideas constituted of simpler ideas. The research program for cognitive science, therefore, became one of showing how the physical world maps onto the mental world, discovering the mental rules that govern ideas and showing how the secondary properties of experiences arise when the physical world has only primary, mechanical, properties.

《理念之路》对认知科学经典立场的修正

The Way of Ideas’ Revisions to the Classic Positions in Cognitive Science

贝克莱对“理念之路”的开发及其根本含义有助于激发对古典认知心理学的修正。

Berkeley’s development of the Way of Ideas—and its radical implications—helped provoke revisions to the classical cognitive psychologies.

我们现在在哪里：思维设计

Where We Are Now: Mind Design

关于人类心智，我们目前有三种描述。从构建心智科学的角度来看，经验主义者提供了最丰富、最详尽的描述。他们提出了一套心理能力，信息通过这些能力从感觉、知觉到记忆进行处理，为科学研究项目开辟了新的前景。

We have three sketches of what the human mind is like. From the perspective of building a science of the mind, the richest and most detailed sketch was offered by empiricists. They proposed a set of mental faculties through which information is processed from sensation to perception to memory, opening vistas for scientific research programs.

现实主义和唯心主义也影响了认知科学。有时，它们作为反对联想主义的抗议运动而有所贡献。现实主义认为，心智并非局限于头脑，而是与世界相连；唯心主义则认为，心智并非世界镜像，而是积极地构建我们对世界的图景。近年来，它们催生了各自的研究项目。现实主义研究我们与世界的身体互动如何塑造我们的认知；唯心主义则推动了认知神经科学中预测加工理论的兴起。

Realism and idealism have influenced cognitive science, too. Sometimes they contributed as protest movements against associationism. In the case of realism, arguing that mind was not locked up in the head but connected to the world; in the case of idealism that mind did not mirror the world, but actively constructed our picture of it. More recently they have spawned their own research programs. In the case of realism by studying how our bodily interaction with the world shapes our cognition; in the case of idealism in the rise of the predictive processing theory in cognitive neuroscience.

带走

Take away

本章的主要收获也是认知心理学的一个重要教训。我们成年人的心智受到两种遗传的影响：一种来自进化，一种来自经验。后者大部分源于我们的文化，而且最重要的是，我们所相信的几乎所有东西都是隐性的，或者说是内隐性的。作为历史延续的参与者，我们通过间接和推理来继承知识和传统。本章试图阐明认知科学家继承下来的关于人性和人类思维的隐性信念和理论。了解这些信念和理论将有助于我们更好地理解认知理论和研究为何会发展到如今的水平，并将极大地阐明不同观点之间的争论。我们往往比争论的双方更能理解这些信念和理论。

The main important takeaway from this chapter is also a major lesson from cognitive psychology. Our adult minds are shaped by two kinds of inheritance, one from evolution and one from experience. Most of the latter is imbibed from our cultures, and, most importantly, almost all of what we believe is tacit, or implicit. We inherit knowledge and traditions by indirection and inference, as participants in the continuing story of history. What this chapter has tried to do is make conscious the implicit beliefs and theories cognitive scientists inherited about human nature and the human mind. Knowing what they are will help us better understand why theory and research on cognition have developed as they have and will throw much light on arguments among rival views. We will often understand them better than the contesting sides.

阅读材料

Readings

本章的主题主要涉及科学哲学、心灵哲学和心理学史。

The topics in this chapter draw mostly from philosophy of science, philosophy of mind, and history of psychology.

更多关于心理学史的内容，请参阅我的《心理学史：从古代到现代》（伦敦：Taylor & Francis/Routledge出版社，2018年）。阅读哲学心理学家的著作，请访问网站https://psychclassics.yorku.ca。关于认识论的入门，请参阅Jennifer Nagel的《知识：非常简短的导论》（牛津：牛津大学出版社，2014年），该书是优秀“非常简短”系列丛书的一部分。关于魔术，请参阅Joshua Jay的《魔术师如何思考》（纽约：Workman出版社，2021年）。

For more on the history of psychology, see my A History of Psychology: From Antiquity to Modernity (London: Taylor & Francis/Routledge, 2018). To read works by philosophical psychologists, see the website https://psychclassics.yorku.ca. For an introduction to epistemology, see Jennifer Nagel, Knowledge: A Very Short Introduction (Oxford: Oxford University Press, 2014), part of the excellent Very Short series. For magic, see Joshua Jay, How Magicians Think (New York: Workman, 2021).

牛津的 Short 系列还包括 P. Okasha 的《科学哲学：非常简短的介绍》（牛津：牛津大学出版社，2002 年）。

Oxford’s Short series also includes P. Okasha, Philosophy of Science: A Very Short Introduction (Oxford: Oxford University Press, 2002).

关于心灵哲学的两本著作分别是P. Mandik的《心灵哲学导论》（纽约：Wiley出版社，2013年）；以及A. Kind的《心灵哲学：基础》（伦敦：Taylor & Francis出版社，2020年）。最佳五本书的列表请访问https://fivebooks.com/best-books/philosophy-mind-keith-frankish/。

Two books on philosophy of mind are P. Mandik, This is Philosophy of Mind: An Introduction (New York: Wiley, 2013); and A. Kind, Philosophy of Mind: The Basics (London: Taylor & Francis, 2020). A list of Best Five is at https://fivebooks.com/best-books/philosophy-mind-keith-frankish/.

电影

Movies

我会推荐一些与认知科学相关的电影。这里我推荐印第安纳大学，它有一个精选认知科学相关电影的网站：https://cogfilms.sitehost.iu.edu。它的起源可以参见B. Motz的《大众电影中的认知科学：认知科学电影索引》，载于《认知科学趋势》第17卷（2013年）：第483-485页。http ://dx.doi.org/10.1016/j.tics.2013.08.002。

I’ll be recommending some movies relevant to cognitive science. Here I point you to Indiana University, which hosts a curated site on films related to cognitive science: https://cogfilms.sitehost.iu.edu. Its origin is described in B. Motz, “Cognitive Science in Popular Film: The Cognitive Science Movie Index,” Trends in Cognitive Sciences 17 (2013): 483–485. http://dx.doi.org/10.1016/j.tics.2013.08.002.

魔法

Magic

最古老的魔术或许是杯球魔术，这在埃及象形文字中得到了证实。佩恩和泰勒非常擅长表演杯球魔术，即使你能看到他们的每一个动作，他们也能表演成功。详情请见：https://www.youtube.com/watch?v =8osRaFTtgHo 。网上有很多教程，你可以选择自己最喜欢的一个。现在有一个魔术科学协会 (SoMA)：https://scienceofmagicassoc.org。

The oldest trick in magic is probably the cups and balls, attested by Egyptian hieroglyphics. Penn and Teller are so good at performing the cups and balls that it works even when you can see every move they make. See https://www.youtube.com/watch?v=8osRaFTtgHo. There are several tutorials on the web, you can pick the one you like best. There is now a Science of Magic Association (SoMA): https://scienceofmagicassoc.org.

第二次世界大战：机器智能的出现

World War II: Emergence of Machine Intelligence

虽然巴贝奇的机器和洛夫莱斯的理念是现代计算机的先驱，但它们并没有催生数字时代、认知科学和人工智能。战争才是真正的催生者（Dyson，2012；Morris，2014）。

While Babbage’s machines and Lovelace’s ideas are forerunners of modern computers, they did not give birth to the digital age or to cognitive science and AI. War did (Dyson, 2012; Morris, 2014).

控制论：自我引导系统、反馈和信息

Cybernetics: Self-Guiding Systems, Feedback, and Information

在讨论机械自动机时，我们忽略了它们与生物体之间的一个重要区别。正如艾达·洛夫莱斯（Ada Lovelace，1842）所指出的，机器只能执行我们设计或编程让它们执行的操作。例如，斯特拉斯堡钟无论周围发生什么，都会报时；在笛卡尔的动物行为反射模型中，固定的刺激会引起固定的反应。但现实中的动物会适应环境，并根据不同的情境对刺激做出不同的反应。最值得注意的是，现实中的动物可以灵活地追寻移动的目标，就像猎豹追逐小羚羊一样。

In discussing mechanical automata, we’ve passed over an important difference between them and living things. As Ada Lovelace (1842) noted, machines can only do what we have built them, or programmed them, to do. The Strasbourg clock, for example, rings the time regardless of what’s going on around it, and in Descartes’ reflex scheme of animal behavior, fixed stimuli give rise to fixed responses. But real animals are attuned to their environment and respond differently to stimuli depending on context. Most notably, real animals can flexibly pursue a moving goal, as when a cheetah chases a dik-dik.

战争或许总是愚蠢的，但19世纪和20世纪初的战争却极其愚蠢且代价高昂。第一次世界大战中的炮火可能投入数十万甚至数百万枚炮弹，但对敌方阵地造成的损害往往很小，因为一旦发射，由于没有精确瞄准而导致炮弹不准确，就会遵循牛顿运动定律，形成弹道。即使在人类炮手的引导下，也需要数百枚防空炮弹才能击落一架缓慢移动的轰炸机。第二次世界大战期间，控制论（该术语的词根源于希腊语，意为“指挥或操控”）的创始人（Rosenblueth、Wiener 和 Bigelow，1943/1966）向人工智能迈出了一大步，他们证明了机器可以有目的地行动，并发明了反馈和信息这两个关键概念（Lloyd，2019）。

War may always be stupid, but the wars of the 19th and early 20th centuries were appallingly stupid and expensive. Artillery barrages in World War I could involve hundreds of thousands, even millions, of shells inflicting often minor damage to enemy positions because, once fired, inaccurately without precise targeting, projectiles go ballistic, following Newton’s laws of motion. Even guided by human gunners, it took many hundreds of anti-aircraft shells to take down a single slow-moving bomber. During World War II, the founders of the field of cybernetics—the term’s root is in Greek words meaning to direct or steer (Rosenblueth, Wiener, & Bigelow, 1943/1966)—took a huge step toward artificial intelligence by showing that machines could behave purposively, inventing the key concepts of feedback and information (Lloyd, 2019).

引导的基本问题是设定目标并采取行动去接近（或有时是避免）它。任何目标导向系统都需要三个组件：一个用于检测目标并确定我们与目标关系的传感器；一个用于改变我们与目标关系的行动系统；以及一个将传感器与行动系统连接起来的控制系统（图 2.1）。

The basic problem of guidance is setting a goal and acting to approach (or, sometimes, avoid) it. Any goal directed system needs three components: A sensor that detects a goal and determines our relation to it, an action system that changes our relation to the goal, and a control system that connects the sensor to the action system (Figure 2.1).

反馈系统的一个例子是恒温器。它有一个简单的目标（将室温保持在 74 华氏度），并调节热泵来实现这一目标。请注意，简单的恒温器本身就是一个与环境相互作用的有机体。当温度偏离其目标时，恒温器会利用热泵来改变环境。当此动作影响环境时，恒温器会停止工作，这以其自身的方式改变温度，直到需要再次启动热泵。正如控制论术语所示，恒温器会不断地将其与环境的关系引导到理想状态。这就是为什么反馈被恰当地称为反馈回路，因为恒温器作用于环境，环境反过来作用于恒温器。

An example of a feedback system is the thermostat. It has a simple goal (keep room temperature at 74 degrees F) and regulates a heat pump to do so. Notice that the simple thermostat is in its own way an organism that mutually interacts with its environment. When temperature is off its target, the thermostat acts to change the environment by using the heat pump. When this action affects the environment, the thermostat ceases to act, which in its own way changes the temperature until the heat pump needs to be called on again. True to the term cybernetics, the thermostat continuously steers its relationship with the environment into an ideal state. That is why feedback is properly called a feedback loop, because the thermostat acts on the environment and the environment acts on the thermostat reciprocally.

这个想法看似简单—— “反馈”^这个词虽然经常被误用，却也如此常见——以至于很难理解它带来的巨大影响。一次简单却毁灭性的打击，它表明，机器并非仅仅是机械装置，至少在某种意义上可以体现智能，能够在特定环境中有效运作。目的并非思维的专属领域，而是可以在机器中实现的。

The idea seems so simple—and the term feedback is so familiar, if often misused³—that it’s hard to understand the huge impact it had. At one simple, yet devastating, blow, it showed that machines need not be mere clockwork but could be intelligent by at least one meaning, behaving effectively in one’s environment. Purpose was not the exclusive realm of mind but could be implemented in a machine.

控制论的另一个关键概念是信息，这在今天也是一个常见的术语，其影响已被视为理所当然。在物理学中，物体通过力和/或粒子的交换相互影响，物理定律将它们作为方程中的变量。然而，在控制论中，物理相互作用的地位低于那些最好被称为“通过思想来做事”的事物（Haugeland，1985）。

The other key concept of cybernetics is information, again a commonplace term today whose impact is taken for granted. In physics, objects affect one another by exchange of forces and/or particles, and physical laws include them as variables in equations. However, in cybernetics, physical interaction takes second place to what can best be labeled regulation by ideas that do things (Haugeland, 1985).

我们稍后会看到，恒温器的工作可以由几种物理上不同的设备完成：读取温度计的人工服务、带有弯曲电线的传统恒温器，或各种现代的计算机化智能恒温器。每种情况都涉及物理原理，因为热是物体原子的活动状态。然而，使恒温器工作的是有关空气温度的信息，而不是空气分子的运动。信息看起来像一个传统的物理概念，因为我们经常会这样对待它。因此，我们可以问：“我们有多少关于俄罗斯军队调动的信息？”就像我们可以问“俄罗斯有多少储备汽油？”一样。但是，如果两辆坦克共用100加仑汽油，它们之间仍然只有100加仑。当它们上路并燃烧汽油时，最终就用完了。换句话说，这就是质能守恒。但是，如果一位将军向所有指挥官分发一份包含俄军动向信息的备忘录，这位将军仍然拥有原始的信息储备；事实上，由于信息共享，将军的部队现在拥有了更多信息。而且，这些信息在战斗中被调用时也不会消失。信息处理发生在迄今为止人们对其代表世界的理念层面上，但只要正确利用，它不仅可以代表世界（如斯多葛学派所见），还可以改变世界（如控制论所表明的那样）。

As we’ll see in a moment, the thermostat’s job can be done by several physically different devices: A human servant reading a thermometer, a traditional thermostat with its bending wire, or a variety of modern computerized smart thermostats. In each case, there are principles of physics involved, because heat is the activity state of an object’s atoms. However, what made the thermostat operate is information about the air temperature, not the movements of the air’s molecules. Information looks like a traditional physical concept because we can often appear to treat it as such. So, we can ask, “How much information do we have about Russian troop movements?” in the same way as we can ask, “How much reserve gasoline do the Russians have?” But if two tanks share 100 gallons of gasoline, there’s still only 100 gallons between them. When they hit the road and burn the gas, eventually there’s none left. In other words, there’s conservation of mass-energy. But if a general circulates a memo to all commanding officers containing the information about Russian movements, the general still has the original store of information; indeed, there’s now more information in the general’s unit, as the information is shared. Nor does the information disappear as it’s called upon during battle. Information processing takes place at a heretofore dimly recognized level of ideas that represent the world, but it can, when harnessed correctly, not just represent the world, as the Stoics saw, but also change the world, as cybernetics showed.

图灵测试与人工智能的可能性

The Turing Test and the Possibility of Artificial Intelligence

1950 年，图灵发表了《计算机器与智能》。这篇文章并没有发表在数学或工程学期刊上，而是发表在第一本英文心理学期刊《心智》上。图灵开篇说道：“我打算思考这个问题：‘机器能思考吗？’”（第 433 页），并提出了模仿游戏，这是一种用人类标准衡量机器智能的测试。他刻意避免对机器或智能进行定义。在一起淫秽案件中，波特·斯图尔特大法官表示，尽管他无法定义色情内容，但他看到时就知道是色情内容。模仿游戏的理念也是如此：智能或许难以被定义，但当我们遇到它时，我们就知道是智能。

In 1950, Turing published “Computing Machinery and Intelligence.” It did not appear in a mathematical or engineering journal but in what had been the first English journal of psychology, Mind. Turing began, “I propose to consider the question, ‘Can machines think?’” (p. 433), and he introduces the imitation game, a test by which to measure machine intelligence against a human standard. He deliberately avoids defining either machine or intelligence. In a case of obscenity, Justice Potter Stewart remarked that although he could not define pornography, he knew it when he saw it. The idea of the imitation game is the same: Intelligence may evade being pinned down, but we know it when we meet it.

最初，游戏由一名询问者（图灵的思想中仍保留着间谍的影子）以及一男一女组成。这名男子和一名女子被单独安置在不同的房间里，并通过电传打字机与询问者匿名互动。询问者的工作是向每位参与者提问，目的是找出哪个是男性；参与者可以根据自己的意愿撒谎或回避问题，以隐藏自己的性别。现在，想象一下女性被一台候选智能机器取代。图灵问道：“当游戏以这种方式进行时，询问者做出错误判断的概率会像在男女之间进行游戏时一样吗？这些问题取代了我们最初的问题：‘机器能思考吗？’”（第434页）。他评论道：

Initially, the game consists of an interrogator (shades of espionage remain in Turing’s thought) and a man and a woman. The man and woman are placed alone in separate rooms and anonymously interact via teletype with the interrogator. The interrogator’s job is to ask questions of each participant with the goal of figuring out which is the man; the participants may lie or evade the questions as they see fit, aiming to hide their sex. Now, imagine the woman is replaced by a candidate intelligent machine. Turing asks, “Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, ‘Can machines think?’” (p. 434). He remarks:

这个新问题的优势在于，它能够在人类的体能和智力之间划出一条相当清晰的界限。没有工程师或化学家声称能够制造出一种与人类皮肤难以区分的材料。或许在未来的某个时候，我们能够做到这一点，但即使假设这项发明能够实现，我们也应该觉得，试图用这种人造血肉来让一台“思考机器”更像人类，也毫无意义^。6

图灵，1950 年，第 434 页

The new problem has the advantage of drawing a fairly sharp line between the physical and the intellectual capacities of a man. No engineer or chemist claims to be able to produce a material which is indistinguishable from the human skin. It is possible that at some time this might be done, but even supposing this invention available we should feel there was little point in trying to make a “thinking machine” more human by dressing it up in such artificial flesh.⁶

Turing, 1950, p. 434

结合我们提出的三种认知架构，图灵揭示了符号系统架构中强大的笛卡尔传统，因为他假设智能行为与动物或人类的身体无关，而只与它们的思维有关。在详细描述了通用数字计算机之后，图灵乐观地认为，到2000年，没有人会认为思考机器是不可能的。

Considering our three architectures of cognition, Turing reveals the powerful Cartesian heritage of the symbol system architecture, in that he assumes intelligent behavior has nothing to do with animal or human bodies, but only with their thoughts. After going on to describe general-purpose digital computers in some detail, Turing is optimistic that, by 2000, no one will regard thinking machines as impossible.

图灵思考了对机器智能可能性的反对意见，包括来自神学甚至超感知觉（ESP）的反对意见。从我们的角度来看，其中一些意见尤为突出。其一是洛夫莱斯夫人（Lady Lovelace）的反对意见，她认为计算机只能执行指令。其二是神经系统是连续的而非数字的机器；然而，由于数字机器是通用计算机，它们可以模仿数字机器。第三种混乱的反对意见是来自意识和艺术情感的反对意见，这些反对意见似乎与计算机的内心世界格格不入。还有一组混乱的论点，来自行为的非形式性，或者用今天的话说，行为的社会嵌入性。这些论点将成为具身认知支持者反对主流认知科学的论据的一部分。图灵最后思考了哪些领域最适合研究机器智能，并做出了惊人的预言（第460页）：“许多人认为像下棋这样非常抽象的活动是最好的。”他的另一个“最佳”想法是制造和抚养一个机器孩子。

Turing considers objections to the possibility of machines being intelligent, including ones from theology and even ESP. From our perspective, a few stand out. One is Lady Lovelace’s objection that computers can only do what they are told. Another is the fact that the nervous system is a continuous rather than digital machine; however, since digital machines are universal computers, they can imitate digital ones. A messy third are objections from consciousness and artistic feelings that seem to have no place in a computer’s heart. Another messy set are arguments from the informality of behavior, or as it would be put today, the socially embedded nature of behavior. These will be part of the case against mainstream cognitive science by partisans of embodied cognition. Turing ends wondering about what fields would be best for investigating machine intelligence and is strikingly prophetic (p. 460): “Many people think that a very abstract activity, like the playing of chess, would be best.” His other “best” idea was building and raising a machine child.

人工智能的灵感源自英国的图灵，但它的名字和资金支持却源于美国。1955年8月31日，由达特茅斯学院数学家约翰·麦卡锡领导的一支即将成为人工智能界巨擘的团队向洛克菲勒基金会提交了“达特茅斯学院人工智能夏季研究项目提案”（McCarthy, Minsky, Rochester, & Shannon, 1955）。21世纪的主导技术就此诞生。这项技术花了一段时间才得以发展。罗切斯特当时担任IBM的研发主管，但股东们担心人工智能往好了说是无用，往坏了说是危险，于是关闭了IBM的人工智能研究。

AI got its inspiration from Turing in England, but it got its name and its funded existence in the United States. On August 31, 1955, a team of soon-to-be AI luminaries led by Dartmouth mathematician John McCarthy submitted to the Rockefeller Foundation, “A proposal for the Dartmouth summer research project on Artificial Intelligence” (McCarthy, Minsky, Rochester, & Shannon, 1955). The dominant technology of the 21st century was born. It took a while to get going. Rochester was head of research and development at IBM, but shareholders, afraid that AI would be at best useless and at worst dangerous, shut down AI research at IBM.

认知科学的创立

Creation of Cognitive Science

图灵、冯·诺依曼和达特茅斯团队提出了机器可以智能的想法，但他们回避了其工作对人类的影响：他们创造了人工智能，但没有创造认知心理学。认知心理学和认知科学的创始人是达特茅斯项目的后来者，他们是赫伯特·西蒙和艾伦·纽厄尔 (1956)。他们在符号系统假说中将图灵的智能是受规则支配的符号操作的想法扩展到人类：智能就是符号操作，无论是在计算机硬件中执行，还是在动物和人类的大脑中进行：“很明显，任何能够执行人类功能的机器都可以通过改变视角转变为行为模型”(Newell & Simon，1956，第 2 页)。

Turing, von Neumann, and the Dartmouth group advanced the idea that machines could be intelligent, but they danced around the implications of their work for human beings: They created AI but not cognitive psychology. The creators of cognitive psychology and thus cognitive science were latecomers to the Dartmouth project, Herbert Simon and Allen Newell (1956). They extended Turing’s idea that intelligence was rule-governed symbol-manipulation to humans in their symbol-system hypothesis: Intelligence just is symbol-manipulation, whether carried out in the hardware of computers or in the brain of animals and people: “It is clear that any machine that can perform human functions can be turned into a model of behavior simply by a change in viewpoint” (Newell & Simon, 1956, p. 2).

解释什么：两种心智概念

What to Explain: Two Concepts of Mind

到目前为止，我们对“心智”一词的使用较为宽泛，但现在我们需要更加精确。大卫·查尔默斯（David Chalmers，1996）曾指出，“心智”的定义和运用方式有两种截然不同的方式：因果性心智概念和意识性心智概念。前者根据心智状态与行为的关系来定义心智状态；后者根据心智状态与意识的关系来定义心智状态。

So far, we’ve used the term “mind” loosely, but now we need to be more precise. David Chalmers (1996) has usefully pointed out that “mind” has been defined and deployed in two quite different ways, the causal concept of mind and the conscious concept of mind. The former defines mental states with regard to their relation to behavior; the latter defines mental states with regard to their relation to consciousness.

让我们从因果概念开始，因为它存在于民间心理学中。假设我因为口渴而翻冰箱，并且我相信里面有一瓶可乐。我的口渴和我的信念相互作用，引发了我的寻找，两者都是有意识的，在图2.2中被称为因果体验。现在假设我在校园里往自动售货机里投入2美元，同时抱怨部门运营预算削减。我相信把钱投入机器就能得到一瓶水，这导致了我的行为，但由于我的心智在其他地方，用笛卡尔的话来说，它目前不在意识中，尽管它可以很容易地被带入意识。因此，这个信念属于图中的认知无意识部分。最后，假设有人问我：“法国的首都是哪里？”我回答：“巴黎。”在我回答的时候，我的脑海中可能浮现出某种意象。在我写作的时候，对很多人来说，脑海中浮现的画面可能是巴黎圣母院大火的场景（2019年4月15日）。然而，这幅画面不太可能让我回答“巴黎是法国的首都”。早在巴黎圣母院起火前，我就已经将答案作为长期记忆中的事实。此外，不同的人可能会报告不同的画面；有些人可能回忆起巴塔克兰剧院的恐怖袭击；另一些人可能回忆起埃菲尔铁塔之旅，或在香榭丽舍大街享用的美味午餐，或学校地理课本上的地图。有些人则完全没有提到任何画面（Beenfeldt，2013；Leahey，2014）。潜意识搜索所需信息可能会（也可能不会）触发记忆，但该记忆的图像在回忆巴黎是法国首都方面并未发挥任何作用。用哲学术语来说，图像是一种附带现象的体验，是一种不会引起行为的意识事件。

Let us start with the causal concept because it’s found in folk psychology. Suppose I’m rooting around in my fridge because I’m thirsty and I believe that there’s a Coke inside. My thirst and my belief interact to cause my search, and both are conscious, located in Figure 2.2 as Causal Experience. Now suppose I’m on campus feeding $2.00 into a vending machine while complaining about cutbacks to my department’s operating budget. My belief that putting money in the machine will get me a bottle of water causes my behavior, but because my mind is elsewhere, in Descartes’ phrase, it’s not currently conscious, though it could be readily brought to consciousness. This belief is thus in the Cognitive Unconscious part of the Figure. Finally, suppose I’m asked, “What is the capital of France?” I reply, “Paris.” It’s possible that as I reply, I have a mental image of some sort. For many people, as I write, the image may be of the cathedral of Notre-Dame burning (April 15, 2019). However, it is highly unlikely that this mental image caused me to reply that Paris is the capital of France. I knew the answer as a fact stored in long-term memory years before Notre-Dame caught fire. Moreover, different people may report different images; some may recall the Bataclan terror attack; others may recall a trip to the Eiffel Tower or a pleasant lunch on the Champs-Élysées or a map from a school geography text. Some people will report no images at all (Beenfeldt, 2013; Leahey, 2014). Unconscious search for the requested information may (or may not) trigger a memory, but the image of that memory played no role in recalling that Paris is the capital of France. In philosophical parlance, the image is an epiphenomenal experience, a conscious event that does not cause behavior.

在笛卡尔之前，像亚里士多德这样的思想家强调心灵的因果性，很少探讨意识。笛卡尔彻底颠覆了古代的优先次序，从本质上将意识发明为哲学关注的一个领域。在认知科学中，厘清这些不同类型的心理事件和过程，尤其是它们之间的关系，将引发争议和困惑。本书将讨论以下问题：

Prior to Descartes, thinkers such as Aristotle emphasized the causal aspect of mind, having little to say about consciousness. Descartes radically reversed ancient priorities, essentially inventing consciousness as a domain of philosophical concern. In cognitive science, becoming clear about these different kinds of mental events and processes, and especially the relations between them, will cause controversy and confusion. Here are issues we will discuss in this book:

• 认知科学应该关注副现象体验吗？由于认知科学主要关注行为的预测和控制，许多心理学家对此持否定态度。人工智能领域的研究人员对此视而不见。
• Should cognitive science concern itself with epiphenomenal experience? Because they are primarily concerned with the prediction and control of behavior, many psychologists have said no. Workers in AI just ignore it.
• 副现象体验的存在暗示着更令人不安的事情，因为它表明意识并非了解行为成因或潜意识心理过程运作的可靠窗口。或许所有意识状态都是副现象，即使是像疼痛这样看似因果的事物。你的脚离火太近，感到疼痛，一边大喊“哎哟！”一边把脚抽回来。疼痛确实感觉像是引发了你的行为，但也可能是一个单一的潜在神经过程导致了疼痛的感受、疼痛的言语表达以及你抽回脚的动作。如果这是真的，那么行为科学可能会完全忽略意识，认为它与预测和控制行为无关。
• The existence of epiphenomenal experience suggests something more troubling because it shows that consciousness is not a reliable window on the causes of behavior or the operations of unconscious mental processes. It might be that all conscious states are epiphenomenal, even something as seemingly causal as a pain. You get your foot too near a fire, feeling pain and yelling ouch! as you pull your foot away. It certainly feels like the pain causes your behaviors, but it could be that a single underlying neural process causes the feeling of pain, its verbal expression, and the withdrawal of your foot. If this is true, then a science of behavior may ignore consciousness altogether as irrelevant to predicting and controlling behavior.
• 无意识状态在多大程度上可供意识报告？即使我的想法只是附带现象（正如刚才提到的），它们也可能是了解无意识过程因果关系的科学方法。在上面的两个例子中，我可以很容易地报告说我认为冰箱里有一瓶可乐，并且可以解释为什么我把绿色的小纸片放在一个钢盒里。另一方面，行为可能是由我们完全没有意识到的信念和过程引起的，而我编造故事来满足我理解自己的需要。例如，左右脑不再交流的患者可能会因刺激他们不会说话的右半球而采取行动。当被问及他们为什么这样做时，这些病人会虚构故事，编造一个看似合理但却是错误的解释。
• To what extent are unconscious states available for conscious report? Even if my thoughts are epiphenomenal, as just suggested, they may be scientifically helpful ways of getting at causal unconscious processes. In my first two examples above, I can easily report that I think there’s a Coke in the fridge and can explain why I’m putting little green pieces of paper in a steel box. On the other hand, it’s possible that behavior might be caused by beliefs and processes of which we are completely unaware, and that I make up tales to satisfy my need to understand myself. For example, patients whose left and right brains no longer communicate can be provoked to action by stimuli presented to their non-speaking right hemisphere. Asked why they acted as they did, such patients confabulate, making up a plausible, but mistaken, explanation.
• 最后，笛卡尔还有一个值得质疑的假设：我们直接了解自己的心理状态和过程，但会根据他人的行为间接地将心理状态和过程归因于他人。因此，我内省地知道自己正在冰箱里找可乐，但如果我看到你在冰箱里翻找，我就会间接地推断出类似的信念来解释它。然而，考虑到上述情况，我们有可能先采取行动，然后再进行合理化，为我们的行为找到理由。这些解释甚至可能是真实的，就像我们解释他人行为时通常所言，但它们并非笛卡尔所定义的特权——我们没有一条通往自身行为原因的特殊渠道。相反，我们观察自己的行为，并构建一个民间心理学解释。我们比他人更了解自己，并非因为我们会内省意识，而是因为我们拥有比他人更大的自身行为样本。如果这是真的，那么研究人员询问受试者的信念和动机就毫无意义了，因为这些信念和动机可以从客观记录的行为中可靠地推断出来。
• Finally, there is Descartes’ questionable assumption that we are directly acquainted with our mental states and processes but attribute mental states and processes to other people indirectly from their behavior. Thus, I introspectively know that I’m looking for a Coke in my fridge, but if I see you searching your fridge, I indirectly infer a similar belief to explain it. However, given the considerations above, it’s possible that we behave first and rationalize afterward, coming up with reasons for our actions. The accounts may even be true, as they usually are when we explain other people’s actions, but they are not privileged in Descartes’ sense—we don’t have a special pipeline into the cause of our behavior. Instead, we observe our actions and construct a folk psychology account of them. We know ourselves better than others not because we introspect consciousness, but because we have a much larger behavior sample of ourselves than we do of them. If this is true, then there is no point in researchers asking subjects about their beliefs and motives because they can be reliably inferred from objectively recorded behavior.

这些问题和争议的结果是，认知心理学对意识的研究和讨论并不多。如果说意识是当今的热门话题，那么人们关注的焦点在于它的神秘性，而不是它的实用性。大脑中的物理过程如何产生（表面上）非物理的意识领域？人工智能领域的研究人员通常对意识视而不见，除非他们偶尔担心机器是否也能拥有意识。

The upshot of these questions and issues is that cognitive psychology has not investigated or discussed consciousness very much. If consciousness is a hot topic today, the focus is on its mystery rather than its utility. How can physical processes in the brain give rise to the (apparently) non-physical realm of consciousness? Workers in AI just ignore consciousness, except on the rare occasions when they worry if machines could become conscious.

现在，我们将研究科学哲学中应用于认知科学的几个主题。

Now, we will look at a few topics in philosophy of science as they apply to cognitive science.

理论正确吗？

Are Theories True?

提出这个问题似乎令人担忧，因为在所有人类机构中，我们都期望科学能带给我们真理，并高呼“相信科学！”的口号。然而，大众对科学的普遍印象过分强调数据收集和从谬误中筛选真理。科学旨在从事实中寻找模式，然后提出解释事实的理论，而这些理论通常包含对我们无法观察到的物体和结构的推论（例如现代量子物理学的多维弦）。它们可能存在，也可能不存在，我们的推论也可能是错误的。哲学家们长期以来一直在争论，并且仍在争论（Leahey，2018），科学应该在多大程度上超越事实，探究隐藏的原因。现实主义者认为应该这样做；反现实主义者则认为不应该这样做。

Asking this question may seem alarming because of all human institutions we expect science to bring us truth, chanting the slogan, “Believe the science!” However, the popular image of science over-emphasizes collection of data and winnowing truths from falsehoods. Science aims to find patterns in facts and then to produce theories that explain the facts, and these theories typically contain inferences to objects and structures we can’t observe (such as the multidimensional strings of modern quantum physics). They may or may not exist, and our inferences can be wrong. Philosophers have long argued, and still argue (Leahey, 2018), about to what degree science ought to go beyond facts to hidden causes. Realists say they should; anti-realists say they shouldn’t.

实在论者追求真理。也就是说，他们认为科学理论应该如实地描述世界的实体和运作，无论它们可见或隐藏。因此，在实在论的物理学论述中，电子是物质的组成部分，无论人们是否能够观察到它们；引力是一种真实的力，即使我们观察到的是引力的效应而不是引力本身。

Realists aim for truth. That is, they say scientific theories should describe the entities and operations of the world, whether visible or hidden, as they really are. Thus, in a realist account of physics, electrons are constituents of matter, whether or not people can observe them, and gravity is a real force, even though we observe the effects of gravity rather than gravity itself.

实在论的反对者是一群乌合之众，通常被称为反实在论者，但我坚持使用一个更古老、更具描述性的名称——工具主义。工具主义者就像亚里士多德与柏拉图之比。柏拉图说，形式存在于人类经验之外，但仍然是真实的。亚里士多德说，实在论对科学的要求过高。科学的职责是为人类提供一个可用的解释——一种工具——以便理解和处理与我们生活的世界息息相关，仅此而已。对自然进行实际的预测和控制就足够了。当牛顿的批评者抱怨他的万有引力定律仅仅假设了一种吸引力，而没有解释它是如何运作的时，他回应说，这样的假设是不必要的；重要的是他的方程式精确地描述和预测了发生的事情，更重要的是，预测了苹果和小行星必然会发生什么。对于反现实主义者来说，科学关乎预测和控制，而非形而上学。

Opponents of realism are a motley crew usually dubbed anti-realists, but I’ll stick with an older, more descriptive name, instrumentalism. Instrumentalists are like Aristotle compared to Plato. Plato said that Forms exist outside human experience but are nevertheless real. Aristotle said that realism demands too much of science. The job of science is to give human beings a usable account—an instrument—with which to grasp and deal with the world we live in, but nothing more. Practical prediction and control of nature are sufficient. When Newton’s critics complained that his law of gravity simply posited an attractive force without explaining how it worked, he replied that such hypotheses were unnecessary; what mattered was that his equations precisely described and predicted what happened, and, more importantly, what must happen to apples and asteroids. For the anti-realist, science is about prediction and control, not metaphysics.

工具主义有严格和宽松两种形式。严格工具主义者是真正的反实在论者，他们认为科学家应该完全避开隐藏的原因。较为自由的工具主义者则允许使用不可观察的实体和过程来解释行为，只要我们不要将它们仅仅视为有用的概念。

Instrumentalism comes in strict and loose forms. Strict instrumentalists are the true anti-realists, maintaining that scientists should eschew hidden causes altogether. More liberal are instrumentalists who allow use of non-observable entities and processes to explain behavior, as long as we don’t take them to be more than useful concepts.

两种相互竞争的理论能同时成立吗？

Can Two Competing Theories Be True at the Same Time?

再次查看图 2.2，并注意神经事件区域。几乎没有认知科学家支持非物理二元论。大多数人认为，无论有意识与否，心理状态和过程都取决于潜在的神经状态和过程。因此，一个关于脑功能生理理论与心理功能认知理论之间关系的问题出现了。自科学革命以来，将两种不同解释层次的理论联系起来的问题就一直存在，并且有两种解决方法：简化和替代。但与心理学家通常的想法相反，简化是好的。对于心理学而言，消除是坏的（Leahey，2008）。

Look again at Figure 2.2 and note the region of Neural Events. Vanishingly few cognitive scientists endorse non-physical dualism. Most view mental states and processes, conscious or not, as depending on underlying neural states and processes. A question therefore arises about the relationship between physiological theories of brain function and cognitive theories of mental function. The problem of relating two theories at different levels of explanation has been around since the Scientific Revolution, and there are two resolutions, reduction and replacement, but contrary to what psychologists usually think, reduction is good. Elimination is bad, for psychology (Leahey, 2008).

在科学革命早期，实验物理学家们建立了你可能在高中就学过的气体定律。这些定律定量地关联了任何气体的三个变量：压力、温度和体积。虽然气体定律是定量的，但它们只是描述性的，无法解释这些定律成立的原因。

Early in the Scientific Revolution, experimentalists established the gas laws you probably learned in high school. These laws quantitatively relate three variables of any gas, its pressure, temperature, and volume. Although the gas laws were quantitative, they were descriptions, offering no explanation of why the laws held.

随后在18世纪，物质的原子理论复兴，为气体定律提供了解释。原子论认为，气体由分子组成，分子分散在空间中，而不是连接成固体。新理论提出，热量仅仅是衡量气体分子运动速度的一个可测量指标；当能量进入气体时，其分子运动加快；当能量离开时，分子运动减慢。反过来，气体中能量的大小会影响气体定律的另外两个变量。如果体积保持不变，例如在高压锅中，气体分子运动加快，从而增加测得的温度（热量）；如果体积发生变化，那么能量的增加会导致气体分子与容纳气体的物体发生更剧烈的碰撞，导致气体体积增大。如果能量离开系统，分子运动减慢，气体的温度和/或体积就会降低。

Then in the 18th century, atomic theory of matter was revived, providing an explanation of the gas laws. Gases, atomism said, were made up of molecules, dispersed in space rather than linked up into a solid. The new theory proposed that heat was simply a measurable index of the rapidity of motion of the molecules composing a gas; when energy entered a gas, its molecules moved more rapidly, and when energy left, its molecules moved more slowly. In turn, the degree of energy in a gas affected the other two variables of the gas laws. If volume was kept constant, as in a pressure cooker, the gas’ molecules moved faster, increasing measured temperature (heat); if volume changed, then increasing energy caused the gas’ molecules to bang harder on whatever contained the gas, causing its size to increase. If energy left the system, molecules moved more slowly and the temperature and/or the volume of the gas decreased.

这个故事是一个将高层次理论简化为低层次理论的例子。经典气体定律适用于宇宙的一部分，即气体领域，并且由于它是定量的，因此可以用来预测和控制……的行为。从高压锅到热气球，各种气体都适用。新原子理论应用于气体，解释了气体定律的工作原理，并将其编织成一幅更广阔的宇宙科学图景。孟德尔假设的基因被分解成DNA片段的过程也经历了类似的过程。

The story is an example of the reduction of a higher-level theory to a lower-level theory. The classical gas laws were true about one part of the universe, the realm of gases, and because it was quantitative, it could be used to predict and control the behavior of gases from pressure cookers to hot air balloons. Applied to gases, the new atomic theory explained why the gas laws worked and wove them into a larger scientific picture of the universe. A similar story holds for the reduction of Mendel’s hypothetical gene to segments of DNA.

然而，早期现代热与物质科学的其他组成部分却遭遇了不同的命运。在原子论复兴之前，热及其相关现象，例如生锈、呼吸作用和电流，都是通过引用诸如燃素和热质之类的流体从一个物体流向另一个物体来解释的。例如，热质从火中流向附近的物体，使其变热。新的原子物质理论将热重新定义为分子运动的程度，使得热质和燃素不再必要，从而将它们从科学中剔除。更普遍地说，当一个旧理论被证明不充分，并被一个新的、更好的理论取代时，就会发生淘汰。

Other components of the early modern science of heat and matter, however, fared differently. Prior to the revival of atomism, heat and related phenomena, such as rusting, respiration, and electric current, were explained by citing the movement of fluids, such as phlogiston and caloric, from one object to another. Thus, for example, caloric flowed from a fire into a nearby object, making it hot. By redefining heat as degree of molecular movement, the new atomic theory of matter made caloric and phlogiston unnecessary, and thus eliminated them from science. More generally, elimination occurs when an old theory is shown to be inadequate and is replaced by a new, better, one.

在认知科学中，一个迫在眉睫的问题是，对大脑过程的生理学解释是否会减少——即解释心理状态和过程如何运作，同时保持它们在解释层面上的有效性——或者将它们作为虚构而消除。

In cognitive science, a looming question is whether physiological accounts of the brain processes will reduce—that is, explain how mental states and processes work while leaving them valid at their level of explanation—or eliminate them as fictions.

临床抑郁症的治疗方式多少体现了其中的利害关系。在20世纪上半叶，抑郁症，尤其是精神分析学家，被认为是一种心理障碍，其成因在于童年早期经历，其治疗必须依靠分析性谈话疗法。超我等心理实体和压抑等心理过程被认为是患者痛苦的解释。然而，在20世纪80年代，人们发现百忧解等药物可以有效治疗抑郁症，而无需解决所谓的童年抑郁症病因，也无需改变患者超我的运作或对压抑的管理。人们开始怀疑超我和燃素一样，都是虚构的，抑郁症和流感一样，是一种生理疾病。抑郁症逐渐被视为一种可以通过药物治疗的脑部疾病。简而言之，一种关于精神状态的生理理论即将取代心理理论，而美国心理学会预见到了这一趋势，开始倡导赋予临床心理学家开处方的特权。

Something of the size of the stakes involved can be seen in how clinical depression is treated. For the first half of the 20th century, depression was regarded, especially by psychoanalysts, as a psychological disorder whose causes lay in early childhood experience and whose treatment had to come from the analytic talking cure. Psychological entities such as the superego and processes such as repression were offered as explanations for patients’ distress. Then, in the 1980s it was found that drugs such as Prozac could effectively treat depression, without addressing the supposed childhood causes of depression nor altering the operations of a patient’s superego or management of repression. The suspicion arose that the superego was as much a fiction as phlogiston, and that depression was as much a physiological disorder as the flu. Depression came to be seen as a brain disorder treatable by drugs. In short, a physiological theory of a mental state was poised to replace psychological ones, and, seeing the writing on the wall, the American Psychological Association began to advocate prescription-writing privileges for clinical psychologists.

虽然职业收入并非认知科学的焦点，但神经科学蚕食认知心理学的威胁却是一个现实问题。例如，考虑休谟提出的联想心理过程。休谟认为联想是一种精神引力，一种使一个想法与另一个想法联系在一起的真实的精神力量。另一方面，从巴甫洛夫时代开始，神经科学家就一直试图探索当动物学习将一个想法（例如铃声）与另一个想法（例如食物的想法）联系起来时，神经系统会发生什么。假设我们发现联想学习总是涉及加强动物神经系统中至少两个神经元之间的突触联系。这或许是一个简化的例子：休谟的邻近定律，“两个同时出现的想法会联想”，可以简化为赫布定律，“同时放电的神经元会连接在一起”。另一方面，我们可以用休谟的赫布定律与此类似。然而，心理联想论者可能会用休谟的相似性定律来反驳。如果我对你说“睡觉”这个词，并让你用脑海中浮现的第一个词来回答，你很可能会用一个意思相近的词来回答，比如“休息”、“小睡”或“打盹”，而且很难看出语义意义如何与神经元或神经递质相对应。

Although professional income is not at issue in cognitive science, neuroscience’s threat to cannibalize cognitive psychology is a live issue. For example, consider the psychological process of association proposed by Hume. Hume thought that association was a sort of mental gravity, a real mental force that caused one idea to bind to another. On the other hand, from the days of Pavlov, neuroscientists have tried to discover what happens in the nervous system when an animal learns to associate one idea (e.g., the sound of a bell) with another idea (e.g., the idea of food). Suppose we discover that associative learning always involves strengthening synaptic bonds between at least two neurons in an animal’s nervous system. This might be an example of reduction: Hume’s law of contiguity, “Two ideas that occur together become associated” reducing to Hebb’s law, “Neurons that fire together wire together.” On the other hand, we might replace Hume’s law with Hebb’s. A psychological associationist might push back, however, with Hume’s law of similarity. If I say to you the word “sleep” and ask you to reply with the first word that comes into your mind, you are likely to reply with a word of similar meaning, such as “rest” or “nap” or “snooze,” and it’s hard to see how semantic meaning could correspond to a neuron or neurotransmitter.

近因：如何回答问题

Proximate Causation: Answering How Questions

英语中有两个词表示提问：How和why。虽然它们有时互换使用，但它们反映了提出和解答谜题和问题的两种不同方式。考虑一起谋杀案。发现一具尸体，犯罪现场调查员和法医开始工作。他们确定死因，即死者的死亡机制。例如，一颗 .38 口径的子弹击碎了死者的大脑，导致其立即丧失重要功能。侦探们补充了因果关系，查明是谁、在哪里以及在什么情况下扣动了扳机。然而，我们仍然想知道谋杀案发生的原因：什么动机导致枪手夺走另一个人的生命？这两个问题是独立的；也许可以完全确定谋杀案是如何发生的，但无法确定谋杀案为什么会发生，因为动机太多或太少。在法律上，人们使用民间心理学来回答“为什么”问题，但科学的进步可能会对其动机归因提出有争议的挑战。在精神错乱辩护中，被告可能会辩称自己患有脑瘤，导致无理的杀人狂怒。因此，杀人动机不明，也缺乏真正的民间心理学依据。此类案件凸显了认知科学和认知神经科学中最令人不安的方面之一：大脑就像物理宇宙一样，是一台遵循自然规律的机器，因此即使在神志清醒的人身上，也根本不存在自由意志。

We have two words in English for asking questions: How and why. Although they are sometimes used interchangeably, they reflect two different ways of posing and answering puzzles and problems. Let’s consider a murder. A body is found dead, and Crime Scene Investigators and Medical Examiners go to work. They determine the cause of death, that is the mechanisms by which the decedent died. So, for example, a 38-caliber bullet shattered the decedent’s brain, causing immediate loss of vital functions. Detectives add to the causal story, working out who pulled the trigger on the gun, where, and in what circumstances. However, we still want to know why the murder occurred: What motive led the shooter to take another human life? These questions are independent; it might be possible to completely determine how the murder happened but not to determine why it happened because there are too many motives or too few. In law, why questions are answered using folk psychology, but advances in science can controversially challenge its attributions of motive. In an insanity defense, it might be argued that a defendant was suffering from a brain tumor causing irrational homicidal rage. Thus, there is no real motive, no real folk psychology why answer to the killing. Cases such as these raise one of the most socially disturbing aspects of cognitive science and cognitive neuroscience: That the brain is a machine following natural laws just as does the physical universe, so there is no free will at all, even among the sane.

第一种近因解释表明，事件的发生是由于自然定律的作用，例如牛顿运动定律和引力定律，或上文提到的气体定律。因此，天文学家通过获取相关天体（例如太阳、地球和彗星）在太空中的当前位置和运动，并运用牛顿定律预测它们未来的路径，从而找到三者的交点，来预测彗星的回归；或者，物理学家通过将体积和压强值代入气体定律，并读出预测的温度来确定高压锅的温度。根据这种观点，解释只是表明，根据物理系统的初始状态加上适用的自然定律的知识，可以预测某个事件（Hempel & Oppenheim，1948）。

The first kind of proximate causal explanation showed that an event happened because of the operation of natural laws, such as Newton’s laws of motion and gravity, or the gas laws mentioned above. Thus, astronomers predict the return of a comet by taking the present position and motion of the relevant bodies in space (e.g., the sun, the earth, and the comet) and use Newton’s laws to predict their future paths to find the intersection of the three; or physicists determine the temperature in a pressure cooker by plugging into the gas laws the values of the volume and pressure and reading off the predicted temperature. In this view, explanation is simply showing that an event could have been predicted from the initial state of the physical system plus knowledge of the applicable natural laws (Hempel & Oppenheim, 1948).

用定律来解释，这种解释在认知科学中很早就出现了。休谟说，他的联想定律支配着精神世界，正如牛顿定律支配着物理世界一样。认知科学中的第一条定量定律是古斯塔夫·西奥多·费希纳（1801-1887）的心理物理学定律，它将感觉的主观强度绘制成引起该感觉的刺激物理强度的对数函数：S = k logR，其中 S 表示感觉，R 表示刺激强度，ka 表示常数，取决于所涉及的感觉。其他早期认知心理学家也试图寻找类似的意识规律，而寻找学习规律正是认知科学在行为阶段的主要目标。

Explanation by laws appeared early in cognitive science. Hume said his laws of association governed the mental universe just as Newton’s laws governed the physical universe. The first quantitative law in cognitive science was Gustav Theodor Fechner’s (1801–1887) law of psychophysics, which plotted the subjective intensity of a sensation as a logarithmic function of the physical strength of the stimulus causing it: S = k log R, where S is the sensation, R the stimulus strength, and k a constant depending on which sense was involved. Other early cognitive psychologists sought to find similar laws governing consciousness and searching for laws of learning was the major goal of cognitive science in its behavioral phase.

然而，当认知科学开始将大脑比作一台运行思维软件的计算机时，一种与医学和工程学相关的不同解释风格应运而生，被称为机制（Bechtel & Abrahamsen，2005；Craver，2007；Glennan，2017）。在这种解释中，你会看到一个物理结构（或者在认知理论中，是计算结构）在执行某些功能，例如泵血的心脏或行走的工匠鸭子。然后，人们会展示如何将该结构分解为协同运作以实现其功能的机制。就心脏而言，它包含各种腔体、动脉和静脉，它们将血液泵送到肺部，并在每次心跳时将血液从身体中回收。就沃康松的鸭子而言，鸭子内部有齿轮、弹簧和滑轮，外部有机械腿，使鸭子行走。

However, when cognitive science began to adopt the metaphor that the brain was a computer running mind as its software, a different style of explanation associated with medicine and engineering came to the fore, called mechanism (Bechtel & Abrahamsen, 2005; Craver, 2007; Glennan, 2017). In this type of explanation, you have a physical (or in cognitive theory, computational) structure that fulfills some function, such as a heart that pumps blood or an artisan’s duck that walks. Then, one shows how the structure can be broken down into mechanisms operating together to accomplish the structure’s function. In the case of the heart there are the various cavities, arteries, and veins that pump blood to the lungs and recover it from the body with each beat of the heart. In the case of Vaucanson’s duck, there are the gears, springs, and pulleys inside the duck and the mechanical legs on the outside, which make the duck walk.

在机械论解释中，几乎不存在对普遍规律的诉求。事实上，任何给定的机制都可能是独一无二的，可能是由某个物种进化而来的，也可能是由某个工程师为某个设备制造的，因此它们独一无二而非普遍适用。机械论解释的关键概念是结构和功能，而不是规律。乍一看似乎很神秘的事情——血液如何在体内流动，或者铁鸭子如何行走——可以通过将大型任务分解成一系列更简单的任务来揭开神秘面纱，每个任务都可以由大型机器内部的更简单的机器完成。设备的持续功能定义了需要解释的内容，而不是像彗星回归这样的单一事件。

In mechanistic explanation, there is little, if any, appeal to universal laws. In fact, any given mechanism might be unique, assembled by evolution in just one species or built by an engineer for just one device, unique rather than universal. The key concepts in mechanistic explanation are structure and function, not laws. What appears to be mysterious at first glance—how does blood move in the body or how does an iron duck walk—is de-mystified by breaking the large task down into a set of simpler tasks, each of which can be accomplished by a simpler machine inside the larger machine. The device’s ongoing function defines what needs to be explained rather than a single event such as a comet’s return.

在医学、工程学和认知科学中，“功能”的概念尤为重要，因为功能可以建立其他学科所没有的规范。我们不会谈论质子功能失常，但我们会谈论心脏或计算机芯片功能失常，当然也会谈论心智功能失常。当我们谈论这些时，我们会从近端问题——血液如何在体内循环或大脑如何产生意识——转向终极问题——血液为什么在体内循环或我为什么有意识？关于心智，我们可以问：如果人类是理性的，为什么我们会做出错误的决定？

In medicine, engineering, and cognitive science, the concept of function is especially important because functions establish norms, which don’t exist in other disciplines. We don’t speak about malfunctioning protons, but we do speak about malfunctioning hearts or computer chips, and we certainly speak about malfunctioning minds. When we do, we pass from proximate questions—how is blood circulated through the body or how does the brain produce consciousness—to ultimate questions—why is blood circulated through the body or why am I conscious? And we can ask with regard to minds, if humans are rational, why do we make bad decisions?

根本原因：回答“为什么”的问题

Ultimate Causes: Answering Why Questions

在认知科学中，“为什么”的问题可以在两个层面上得到解答：个体生物体层面和物种层面。在每种情况下，“为什么”的问题都考察生物体或物种的战略行为与其目标之间的关系。由于战略是由目标导向的，因此可以提出一些在解释非生物行为时不会出现的规范性问题。我们可以用牛顿定律来解释行星如何围绕其太阳旋转，但行星没有目标；我们不能问一个这颗行星在执行其轨道时表现得好或坏，因为它实际上并没有试图完成任何事情。

Why questions can be answered at two levels in cognitive science, the level of the individual organism and the level of a species. In each case, why questions examine the relationship between an organism’s or species’ strategic behaviors and its goals. Because strategies are goal-directed, it is possible to ask normative questions that do not arise in accounting for the behavior of non-living things. We can use Newton’s laws to explain how a planet revolves around its sun, but planets do not have goals; we cannot ask if a planet is doing a good job or a bad job in executing its orbits because it’s not actually trying to accomplish anything.

生物也是如此。试想一下我们假设的凶手。从道德规范的角度来看，谋杀当然是错误的，但考虑到凶手的目的，我们可以思考他的行为是否经过精心设计，是否有利于实现其策略；事实上，这是评估凶杀现场和抓捕凶手的关键。例如，如果受害者的钱包不见了，我们假设这起谋杀是抢劫计划的一部分。如果钱包还在，那么犯罪可能是一时冲动的冲动表现。

Living things are. Consider our hypothetical murderer. From a moral normative perspective, of course, murder is wrong, but given the murderer’s goal, we can ask if his or her behavior is well chosen to further his or her strategy; indeed, doing so is a key part of evaluating a murder scene and catching the killer. For example, if a victim’s wallet is missing, we hypothesize that the murder was part of a robbery plan. If the wallet remains, the crime may have been a spontaneous expression of passion.

自希腊时代起，思想家们就认为人类是所有生物中特别理性的，甚至是独一无二的理性者，通常会选择实现目标的最佳途径。这种关于人类内在理性的思想，在启蒙运动时期将科学思维扩展到人类社会中至关重要（Smith，2019）。然而，认知科学和社会科学领域的大量研究对这一假设泼了一盆冷水，并对经济和公共政策产生了深远的影响。

From the time of the Greeks, thinkers have assumed that human beings are especially rational, even uniquely rational, among living creatures, typically choosing the best means of achieving their goals, and this idea of intrinsic human rationality was central to the Enlightenment’s extension of scientific thinking to human society (Smith, 2019). However, a great deal of research in cognitive and social science has thrown cold water on this assumption, with deep implications for economic and public policy.

达尔文进化论的应用，对生物普遍的战略行为提出了类似的问题。生存竞争赋予每种生物最终目标：繁衍其基因，将其贡献给其所属物种的持续基因库。因此，正如我们可以探究工人的幸福退休目标是否与其投资策略相符（Schrager，2019），我们也可以探究松鼠的坚果采集和储藏策略是否能带来繁殖成功。即使是树木也会“理性地”追求目标，以最佳的方式生长和伸展叶片，使其能够更好地接受阳光照射。松鼠和树木并没有有意识地提前计划，但它们的行为却常常像在做计划一样。

Application of Darwin’s theory of evolution raises parallel questions for the strategic behavior of living things in general. The struggle for existence sets for each creature the ultimate goal of reproducing its genes, contributing them to an ongoing gene pool of the species of which it is a member. Thus, just as we can ask if a worker’s goal of a happy retirement is well served by his or her investment strategies (Schrager, 2019), we can ask if a squirrel’s nut foraging and caching strategies will lead to reproductive success. Even trees pursue goals “rationally,” growing and spreading their leaves in patterns optimal to exposing them to the sun. Squirrels and trees don’t consciously plan ahead, but they frequently act as if they do.

自然心智设计：认知心理学

Natural Mind Design: Cognitive Psychology

人类长期以来一直渴望飞翔，从代达罗斯和伊卡洛斯的古典时代到文艺复兴时期，发明家们一直从鸟类及其羽毛中寻找灵感^。7 鸟类提供了工程师们所谓的“存在性证明”。如果你想建造某样东西，看看它是否曾经被大自然或其他工程师完成过。鸟类是重于空气的飞行的自然存在证明，因此飞行是可能的。因此，实现工程目标的一种策略是逆向工程：从一个能完成你想要的功能的设备开始逆向设计，然后模仿或复制它（Engard，2016）。

Humans have long wanted to fly, and from the classical days of Daedalus and Icarus and through the Renaissance, inventors looked to birds and their feathers for inspiration.⁷ Birds provide what engineers call an existence proof. If you want to build something, see if it’s been done before, either by nature or by other engineers. Birds are a natural existence proof of heavier than air flight, so flying is possible. One strategy to achieve an engineering goal is thus reverse engineering: Work backward from a device that does what you want and imitate or replicate it (Engard, 2016).

这里有一个真实的例子（Hackman，2004）。20 世纪 80 年代初，英特尔公司垄断了 IBM 架构家用电脑所用 8088 处理器的生产。由于 8088 芯片受专利和版权保护，任何人打开芯片复制硬件或下载软件都是违法的。一家之前获得英特尔授权生产芯片的公司——超微半导体公司 (AMD)——决定尝试逆向工程。AMD 的具体做法极其复杂，也引发了诸多诉讼，但基本上，他们无需窥探 8088 芯片内部结构，就构建了一个芯片和模拟其输入输出功能的软件。从功能上（如果不是从内部机制来看），AMD 芯片就是8088 芯片，AMD 可以以低于英特尔 8088 芯片的价格将其出售给 PC 制造商，从而结束了英特尔的垄断。英特尔在后来的一些芯片上也重复了这一过程，例如20世纪90年代的80386和80486。最终，AMD推出了与之竞争的芯片架构，但由于各种原因，其业务走向衰落（Singer，2012）。

Here’s a real-life example (Hackman, 2004). In the early 1980s, Intel Corporation exercised a monopoly on the manufacture of the 8088-processor used to power the IBM architecture home computers. Because the 8088 was protected by patents and copyrights, it was illegal for anyone to simply open the chip to copy its hardware or to download its software. A company that had previously made chips under license from Intel, Advanced Micro Devices (AMD), decided to try reverse engineering. The details of how AMD did it were fearsomely complex and much litigated, but basically, without peeking into the 8088, they constructed a chip and software that emulated its input-output functions. Functionally, if not internally in its mechanisms, the AMD chip was an 8088 chip, and AMD could sell it to PC manufacturers for less than Intel sold the 8088, ending Intel’s monopoly. They repeated the process with some of Intel’s later chips, such as the 80386 and 80486 in the 1990s. Eventually, AMD offered rival chip architectures, but for various reasons its business went into decline (Singer, 2012).

认知心理学的目标是科学地研究和解释人类智力。尽管人类心智/大脑不受专利和版权保护，但认知心理学家发现自己的处境与AMD大致相同。我们无法内省无意识的心理结构和过程，例如联想；尽管神经科学取得了进展，我们也无法细致地了解大脑的运作。因此，认知心理学家进行逆向工程。在实验中，我们操纵人类心智的输入（用心理学术语来说，即自变量），记录由此产生的行为输出（因变量），并提出关于将输入转换为输出的功能和结构的理论。

The goal of cognitive psychology is to scientifically investigate and explain human intelligence. Although the human mind/brain is not protected by patents and copyrights, cognitive psychologists find themselves in much the same position as AMD. We cannot introspect unconscious mental structures and processes, such as association, and, despite advances in neuroscience, we cannot see the operations of the brain in fine-grained detail. Thus, cognitive psychologists practice reverse engineering. In experiments, we manipulate the inputs to the human mind (in psychology’s jargon, independent variables), record the resulting behavioral outputs (dependent variables), and propose theories about the functions and structures that convert the inputs to the outputs.

人工智能设计：两种人工智能

Artificial Mind Design: Two Kinds of AI

人工智能也始于逆向工程，但目标不同。在认知心理学中，目标是科学性的，运用逆向工程策略来解释人类智能的运作方式。在人工智能中，目标是实用性的，运用逆向工程策略来构建可运行的自主设备，例如自动驾驶汽车。人工智能比认知心理学受到的限制更少，因为尽管它最初研究人类是为了证明智能的存在，但工程师们发现，智能可以通过非人类的方式实现。想要制造飞行器的工程师最初研究的是鸟类和羽毛，但人类的飞行器没有羽毛。

Artificial intelligence also begins with reverse engineering, but with a different goal. In cognitive psychology, the goal is scientific, using the strategy of reverse engineering to explain how human intelligence works. In artificial intelligence, the goal is practical, using the strategy of reverse engineering to construct working autonomous devices such as self-driving cars. AI is less constrained than cognitive psychology because, although it began by studying humans as the existence proof of intelligence, engineers have found that intelligence may be achieved in non-human ways. Engineers who wanted to build flying machines began by studying birds and feathers, but human aircraft don’t have feathers.

认知心理学与人工智能之间的关系错综复杂。区分两种类型的人工智能至关重要；这种区别可能比较模糊，但务必牢记。维度的一端是计算机模拟，另一端是纯人工智能。

The relationship between cognitive psychology and artificial intelligence is a complex one. It is useful to distinguish two kinds of AI; the distinction can be fuzzy but it is important to keep in mind. The first end of the dimension is computer simulation, and the other is pure AI.

用国际象棋来说明这种差异是一个很好的例子。国际象棋被称为人工智能中的果蝇，因为它比任何其他人类任务都得到了更广泛、更深入的研究和逆向工程，或许是因为它似乎是纯粹、理性、无私智慧的巅峰。此外，国际象棋还有一些特殊之处，使其从程序编写的角度来看颇具吸引力。它的规则明确而正式，战斗结果和胜利条件定义明确。它在有限的、明确界定的场地上进行。它是一种信息完全对称的游戏：双方都知道所有棋子的身份和位置，而不仅仅是自己的棋子。最后，它是一种古老的游戏，几个世纪以来一直被人们玩、讨论、分析和记录。至少从表面上看，与其他智力活动相比，关于国际象棋的一切已知信息都已为人所知，这使得在人工智能计算机中编码国际象棋智能成为可能。

A good way to illustrate the difference is with chess. Chess has been called the drosophila of AI because it has been investigated and reverse-engineered more extensively and intensively than any other human task, perhaps because it seems to be a sort of pinnacle of pure, rational, disinterested intelligence. In addition, chess has special features that make it attractive from a program-writing perspective. Its rules are explicit and formal, and outcomes of combat and conditions of victory are well defined. It is played on a limited, well-defined playing field. It is a game of complete and symmetrical information: both players know the identity and placements of all pieces, not just theirs. Finally, it’s an ancient game played, discussed, analyzed, and recorded for centuries. It appears, on the surface at least, that, compared to other intellectual endeavors, everything that can be known about chess is known, making it possible to encode chess intelligence in an artificially intelligent computer.

第一批国际象棋程序，例如AT&T贝尔实验室的Belle，紧密遵循认知心理学的逆向工程策略，将人类的国际象棋策略重写为计算机代码。简而言之，Belle是对人类国际象棋水平的模拟，尽管工程师的目标是提高国际象棋水平，而不是对人类思维进行科学解释。然而，随着计算机计算能力的爆炸式增长，模拟策略变得不再实用。人类，即使是国际象棋大师，也常常不善于详细阐述自己的战术，而且他们受到记忆力和思维速度的限制。而计算机则能够每秒计算数百甚至数千个位置。在人工智能领域，国际象棋的蛮力攻击方法盛行。工程师们放弃了模仿人类，转而依靠计算机纯粹的数字运算能力来击败速度较慢但更狡猾的人类。简而言之，国际象棋人工智能放弃了计算机模拟，转而追求纯粹的人工智能。其目标是胜利，而不是与人类相似。最近，在征服被许多人认为比国际象棋更复杂的中国围棋游戏时，人工智能工程师完全跳过了模拟，而是追求纯粹的人工智能策略，让他们的人工智能自学——他们根本没有编写围棋程序——并掌握游戏。

The first chess programs, such as AT&T Bell Labs’ Belle, hewed closely to the reverse-engineering strategy of cognitive psychology, rewriting human chess strategies as computer code. In short, Belle was a simulation of human chess mastery, even though the engineers’ goal was good chess-playing rather than a scientific explanation of the human mind. However, as the computing power of computers exploded, the simulation strategy became less useful. People, even grandmasters of chess, were often poor at articulating their tactics in detail, and they were subject to limitations of memory and speed of thought. Computers, on the other hand, could calculate hundreds, then thousands of positions a second. The brute-force approach to chess prevailed in AI. Engineers gave up on imitating humans and relied on the sheer number-crunching ability of computers to defeat slower, if more cunning, humans. In short, chess AI abandoned computer simulation for pure artificial intelligence. Victory, not human likeness, was the goal. More recently, in conquering the Chinese game Go, considered by many to be more complex than chess, AI engineers skipped simulation altogether and pursued a pure AI strategy, allowing their AI to teach itself—they wrote no Go program at all—and master the game.

计算机模拟（和逆向工程认知心理学）之间的区别看似微不足道，但事实并非如此。当人工智能遇到自然愚蠢时，很容易从构建模拟滑向构建纯人工智能，同时仍将其误认为是人类认知的模型（McDermott，1976）。人们可能从一个非常简单的目标开始，例如建立一个人类如何玩井字游戏的模型，从井字游戏推理理论开始，并编写代码来执行该理论。然后程序崩溃了。你在代码中找到一个错误并修复它，安装程序员所说的hack或kluge。程序再次崩溃。你重复这种乏味的辛劳，直到程序顺利运行。

The distinction between computer simulation (and reverse-engineered cognitive psychology) may seem trivial, but it is not. It is easy to slide from building a simulation to building a pure AI, while still mistaking it for a model of human cognition, when artificial intelligence meets natural stupidity (McDermott, 1976). One may start with a very simple goal, such as building a model of how humans play tic-tac-toe, beginning with a theory of tic-tac-toe reasoning and writing code to execute the theory. And then the program crashes. You locate a mistake in your code and fix it, installing what coders call a hack or a kluge. The program crashes again. You iterate this tedious toil until the program runs smoothly.

但它充满了黑客攻击。它仍然是你理论的计算版本，还是已经变成了别的东西，某种人造的东西？认知心理学中任何理论的计算实例都可能被问到同样的问题。科学家们最初的目标是创建一个心理过程的计算机模拟，比如解决问题，将他们的理论转化为可运行的代码，但为了让程序运行，他们发现必须安装一些不属于他们心理学理论的“临时解决方案”，但这些方案确实能让程序运行。结果是，随着他们修改程序，他们从计算机模拟转向了纯粹的人工智能，逐渐创建了一个可以工作的逆向工程设备，但它可能不再能够计算人类思维的运作，而​​这恰恰是他们最初的目标。

But it’s full of hacks. Is it still a computational version of your theory or has it become something else, something artificial? The same question may be asked of a computational instantiation of any theory in cognitive psychology. Scientists may begin with the goal of creating a computer simulation of a mental process, such as problem-solving, translating their theory into working code, but to get the program to run they find they must install kluges that are not part of their psychological theory, but which do allow the program to run. The upshot is that as they modify the program they drift from computer simulation to pure artificial intelligence, incrementally creating a reverse-engineered device that works but that may no longer compute the operations of the human mind, their original goal.

我们会发现，理解成功的人工智能哪些部分能够阐明人类智能的某些方面，哪些部分仅仅是人工智能的魔术，是深刻而持久的。当我们谈到当今应用人工智能的最大趋势——深度学习、Facebook 背后的算法、自动驾驶汽车以及计算机医生时，这一点尤为重要。有了深度学习，程序可以自行编码，没有人知道它们内部是如何运作的。

We will find that coming to understand which parts of successful AI illuminate aspects of human intelligence and which are just AI magic tricks is deep and durable. It will be especially important when we come to today’s biggest trend in applied AI, deep learning, the algorithms behind Facebook, self-driving cars, and computer physicians. With deep learning, programs code themselves and no one knows how they work inside.

理论策略：大卫·马尔的分析层次

Strategy for Theory: David Marr’s Levels of Analysis

试图仅通过研究神经元来理解感知，就像试图仅通过研究羽毛来理解鸟类飞行一样：根本行不通。为了研究鸟类飞行，我们必须了解空气动力学：只有这样，羽毛的结构和鸟类翅膀的不同形状才有意义。

马尔，1982年

Trying to understand perception by studying only neurons is like trying to understand bird flight by studying only feathers: it just cannot be done. In order to study bird flight we have to understand aerodynamics: only then do the structure of the feathers and the different shapes of bird wings make sense.

Marr, 1982

大卫·马尔 (1945-1980) 是一位英国认知神经科学家，他在其巨著《愿景》出版后不久因恶性脑瘤不幸逝世。 《愿景》是对贝克莱老问题的开创性、革命性探究，即当深度知觉始于投射在视网膜上的二维图像时，我们如何看待三维空间。然而，更重要的是，马尔为认知科学贡献了一个框架，阐明了逆向工程的过程。上面的段落通常浓缩为“研究飞行，而不是羽毛”，告诉我们要通过仔细分析生物体或设备必须解决的问题来开始逆向工程，而不是迷失在其机制的细节中。古典哲学家认为（有些人继续认为 [Fodor, 2000]）拥有思想是为了发现真理，但认知科学的核心思想是，思想是达尔文生存的工具的集合（Boyer, 2018）。

David Marr (1945–1980) was a British cognitive neuroscientist who tragically died from an aggressive brain tumor shortly after the publication of his great book, Vision. Vision was a pioneering, revolutionary inquiry into Berkeley’s old problem, how we see 3-D space when depth perception begins with a 2-D image cast on the retina. More importantly, however, Marr contributed to cognitive science a framework that illuminated the process of reverse engineering. The passage above, often condensed to “study flight, not feathers,” tells us to begin reverse engineering by carefully analyzing the problem the organism or device must solve rather than getting lost in the details of their mechanisms. Classical philosophers assumed (and some continue to assume [Fodor, 2000]) that having a mind was for discovering Truth, but a central idea of cognitive science is that mind is instead a collection of instruments for Darwinian survival (Boyer, 2018).

Marr 提出，认知科学中的逆向工程应该在三个层面进行：认知层面，它回答这个问题：一个给定的心理过程解决了什么达尔文问题？算法层面，它回答这个问题：该过程在解决问题时遵循什么规则？实施层面，它回答这个问题：大脑和神经系统如何执行这些规则？（图 2.3）。

Marr proposed that reverse engineering in cognitive science should proceed at three levels, the cognitive level, which answers the question, what Darwinian problem does a given mental process address?, the algorithm level, which answers the question, what rules does the process follow in addressing its problem?, and the implementation level, which answers the question, how are the rules executed by the brain and nervous system? (Figure 2.3).

至关重要的是要记住，马尔的分析水平是像认知科学家一样思考的工具。学生们经常将它们与认知心理学中一个听起来类似的框架——处理水平——相混淆，后者的想法是，根据实验中的指示，人们根据浅层物理特征（例如，听列表中押韵的单词，如big-dig）或深层语义（例如，听意思相同的单词，如big-large）来处理刺激。马尔的层次不是心智的信息处理组成部分。它们构成了实践认知科学的策略。

It is crucially important to keep in mind that Marr’s levels of analysis are tools for thinking like a cognitive scientist. Students often confuse them with a similar-sounding framework in cognitive psychology called levels of processing, in which the idea is that depending on instructions in an experiment, people process stimuli in terms of their shallow physical features (e.g., listen for words in a list that rhyme such as big-dig) or their deeper semantic meanings (e.g., listen for words meaning the same thing, such as big-large). Marr’s levels are not information processing components of the mind. They comprise a strategy for practicing cognitive science.

让我们用一个物理工程的例子——不起眼的恒温器——来看一下每个层面。幸运的是，我们不必担心热泵的热力学原理来调节它。8^我们的系统应该追求的目标很简单，就是保持建筑物温度均匀，不太热也不太冷。传统的哲学目标诸如获取知识！或做正确的事！等目标过于模糊，无法用于对心智进行逆向工程，因此必须将其细分为一系列更小的目标。

Let’s look at each level by using a physical engineering example, the humble thermostat. Fortunately, we don’t have to worry about the thermodynamics of a heat pump to regulate one.⁸ The goal our system should pursue is easily stated, to maintain a building at an even temperature, not too hot and not too cold. Traditional philosophical goals such as achieve knowledge! or do the right thing! are too vague to be useful for reverse engineering the mind and will have to be subdivided into sets of smaller goals.

为了实现我们的目标，我们需要三样东西，其中之一就是热泵，我们已经有了。接下来假设现在是冬天，我们只使用暖风。第二样东西，我们需要一种测量温度的方法，这样我们就能知道什么时候温度合适，以及一种通过测量来控制热泵的开关的方法。

To reach our goal we need three things, one of which, the heat pump, we already have. In what follows assume it’s winter and we use only heated air. The second thing we need is a way of measuring the temperature, so we’ll know when it’s right, and a way of turning the heat pump on and off controlled by the measurement.

但是按照 Marr 的说法，我们不会直接着手制造一台机器。相反，我们接下来会看看是否存在一组规则或算法，当设备遵循这些规则或算法时，它们会控制热泵。让我们首先问一下，一个人会怎么做才能将某个区域保持在目标温度，例如 74 华氏度。要测量温度，请将温度计放在房间内方便的位置。然后定期观察温度。当温度低于 74 华氏度时，就下楼打开热泵（或者像以前一样，启动燃煤锅炉）。回到温度计旁，等待温度升到 74 华氏度。当温度升到 74 华氏度时，就下楼关掉热泵或锅炉。重复。我们可以将这些规则重写为计算：

But following Marr, we don’t jump right in to build a machine. Instead, we next see if there is a set of rules, algorithms, that, when followed by a device, will control the heat pump. Let’s begin by asking what a person would do to keep an area at a target temperature of, for example 74 degrees Fahrenheit. To measure the temperature, place a thermometer at a convenient spot in the room. Then regularly observe the temperature. When it gets below 74, go downstairs and turn on the heat pump (or in the old days, stoke the coal-fired boiler). Go back to the thermometer and wait for the temperature to rise to 74. When it does, go downstairs and turn off the heat pump or boiler. Repeat. We can rewrite these rules as computations:

• 10.输入温度作为变量 T。
• 10. Input temperature as variable T.
• 20. 如果 T = 74，则不执行任何操作。
• 20. If T = 74, do nothing.
• 30. 如果 T < 74，则打开热泵或锅炉。
• 30. If T < 74, turn on heat pump or boiler.
• 40. 如果 T > 74，关闭热泵或锅炉。
• 40. If T > 74, turn off heat pump or boiler.
• 50. 转至 10。
• 50. Go to 10.

遵循或受这些算法控制的系统将是智能的，即自适应的，尽管仅限于其狭窄的专业领域。

A system that follows or is governed by these algorithms will be intelligent, i.e., adaptive, albeit only in its narrow domain of expertise.

现在到了实际工程阶段，建造一个装置——恒温器——来调节真实房间的温度。这也是认知架构、认知科学中的实在论以及关于简化和替代的争论变得严肃而充满争议的时刻。

Now comes the practical engineering, building a device—a thermostat—that regulates the temperature in real rooms. This is also the point at which debates among the architectures of cognition, about realism in cognitive science, and about reduction and replacement become serious and contentious.

让我们先来看看数字时代之前制造的老式恒温器。如果你有机会接触到这样的恒温器，可以揭开前盖，看看它的内部工作原理，但这并没有什么意义。它的关键部件可能几乎看不见，但它是一个大约1/16英寸厚、1/8英寸宽、只有几英寸长的小金属条，被卷成一个直径约1到2英寸的线圈，安装在恒温器的背板上。它由两根纵向延伸的细金属条组成，每根金属条对温度变化的响应不同。一根金属条响应迅速，另一根响应缓慢。因此，随着温度的变化，金属条会略微向左或向右弯曲，并可以放置在接触电触点的位置，从而接通或断开电路。当你设定所需的温度（74华氏度）时，金属条会移动，以免接触触点。然而，随着空间冷却，金属条会向触点弯曲，直到温度达到约72华氏度时才接触触点。接触打开了加热器，空间变暖。杆放松直到断开接触，关闭加热器，让房间冷却。该过程不断重复，使房间温度保持在 74 华氏度（约 23 摄氏度），不会太热也不会太冷。请注意，算法层规则描述了恒温器的行为，但它们并不包含在内，恒温器本身也没有真正执行这些规则。它受物理定律的支配。

Let’s begin by looking at old-fashioned thermostats built before the digital age. If you have access to one, you can pull off the front cover and see its inner workings, which won’t mean much. The key component may be barely visible, but it’s a small metal bar about 1/16 of an inch thick and 1/8 of an inch wide, and just a few inches in length, formed into a coil about an inch or two across, mounted on the backplate of the thermostat. It’s made of two thinner bars running lengthwise, each of which responds differently to changes in temperature. One bar responds quickly; the other slowly. As a result, the bar will bend slightly left or right as the temperature changes and can be placed so as to touch an electrical contact, making or breaking an electrical circuit. When you set your desired temperature, 74 degrees F, the bar is moved so as to not touch the contact. However, as the space cools, the bar bends to the contact until it makes contact at about 72 degrees. Contact turns on the heater, and the space warms. The bar relaxes until it breaks contact, turning off the heater, allowing the space to cool. The process iterates, keeping the space at about 74 degrees, not too hot and not too cool. Note for now that the algorithm level rules describe the behavior of the thermostat, but they are not in it, nor is the thermostat actually executing the rules. It is governed by the laws of physics.

现在，让我们来看看现代物联网智能恒温器。揭开前面板，你对它的工作原理了解得比老式恒温器更少，因为你会看到密封的计算机芯片。在智能恒温器中，算法被编程到恒温器中并控制它。在大型现代建筑中，例如教室，温度控制更加复杂，每个房间都有自己的温度传感器，并向中央计算机（例如机械室）报告，中央计算机根据规则对建筑物内每个房间和大厅的空气流量进行不同的调节。

Now, let’s look at a modern Internet-of-things smart thermostat. Pull off the front plate, and you’ll learn even less about how it works than the old thermostat because you’ll see sealed computer chips. In a smart thermostat, the algorithms are programmed into the thermostat and control it. In large modern buildings, such as classroom buildings, temperature control is even more sophisticated, with each room having its own temperature sensor reporting to a central computer (look around for a mechanical room), which uses rules to regulate the flow of air differently to each room and hall in the building.

这个关于非常简单、边缘智能的人工智能的故事包含了一些关于认知科学的要点。

This story about a so-simple, marginally intelligent bit of AI contains some essential points about cognitive science.

• 多重可实现性。这是最广泛、最重要的一点，其他要点都以此为依据。不止一种工作装置能够实现马尔认知层面设定的目标。
• Multiple Realizability. This is the broadest and most important point, from which the others depend. More than one working device can accomplish the goal set out at Marr’s cognitive level.
• 实在论与工具主义。由于多重可实现性的存在，人们开始质疑在认知科学中什么才算解释。描述系统行为的算法是否构成对该行为的解释？我们可以利用规则预测系统在给定任何输入的情况下会做什么，也可以控制系统（这也是我们最初构建它的原因），但旧式恒温器实际上并不遵循规则，因为它们没有经过编程；而智能恒温器则遵循。从科学哲学的工具主义视角来看，了解算法可以解释这两种恒温器的行为；但从实在论视角来看，这只能解释智能恒温器的行为。
• Realism and Instrumentalism. Because of multiple realizability, questions arise about what counts as an explanation in cognitive science. Do the algorithms that describe a system’s behavior constitute an explanation of that behavior? We can use the rules to predict what the system will do given any input, and we can also control the system (which is why we built it to begin with), but old thermostats don’t actually follow the rules because they are not programmed; smart thermostats do. From the instrumental perspective in philosophy of science, knowing the algorithms is an explanation of the behavior of both thermostats, but from a realist perspective, it’s an explanation only of smart thermostats’ behavior.
• 简化与消除。大多数认知科学家采用多重可实现性来避免简化和/或消除心智计算理论（Shouten & de Jong，2007；Leahey，2008）。如果相同的计算理论适用于截然不同的物理设备（在本例中指我们的三种恒温器），那么将一个理论简化为多个理论除了增加复杂性之外，没有任何好处。例如，你不需要知道微软Word如何在英特尔或AMD芯片上工作，你只需要知道它遵循的规则即可使用它。
• Reduction and Elimination. Multiple realizability is taken by most cognitive scientists to fend off reduction and/or elimination of computational theories of the mind (Shouten, & de Jong, 2007; Leahey, 2008). If the same computational theory applies across radically different physical devices (in this case, our three types of thermostats), then nothing is gained but complexity in reducing that one theory to several. You don’t need to know how, say, MS Word works in an Intel or AMD chip to use it, you only need to know the rules it follows.
• 认知架构
  • 神经网络。多重可实现性也意味着认知科学不必局限于单一的解释方法。过去人们认为，认知科学中的符号系统和神经网络方法是相互竞争的。现在人们普遍认为，有些心理过程最好被认为是符号化的，而另一些则是亚符号化的，即联结主义的。
  • 具身认知。另一方面，许多具身认知的倡导者则更加固执己见（Chemero，2009）。他们认为，非计算性的像老式恒温器这样的系统，不应该用计算理论来解释，而应该用设备的物理理论来解释。对他们来说，在这种情况下，算法层面只是一种说法。只有物理实现才能提供真正的解释。
• The Architectures of Cognition
  • Neural Nets. Multiple realizability also means that cognitive science need not limit itself to a single approach to explanation. It used to be thought that the symbol system and neural net approaches in cognitive science were rivals. It’s now accepted that some mental processes are best thought of as symbolic, whereas others are sub-symbolic, that is, connectionist.
  • Embodied Cognition. On the other hand, many advocates of embodied cognition are more obdurate (Chemero, 2009). They say that a non-computational system, such as the old-fashioned thermostat, should not be explained by a computational theory but by a physical theory of the device. For them, in such cases the algorithm level is just a way of talking. Only physical implementation provides real explanations.

研究策略：推断最佳解释

Strategy for Research: Inference to the Best Explanation

马尔的分析层次为我们提供了逆向工程思维和构建心理结构与功能理论的工具。现在的科学问题是如何检验这些理论。

Marr’s levels of analysis have given us tools for reverse engineering the mind and constructing theories of mental structure and function. The scientific question now becomes testing these theories.

与笛卡尔的观点相反，大多数心理过程发生在意识之外，无法被其拥有者观察或内省（Beenfeld，2013；Leahey，2014）。康德接受了这一观点，并得出结论：心理学不可能成为一门科学，因为其研究对象无法通过经验观察。然而，事实证明，即使我们无法看到心理过程，也能够推断出它。这种技术被称为最佳解释推理，有趣的是，它更多地基于规律而非科学。

Contrary to Descartes, most mental processes take place outside awareness and cannot be observed, or introspected, by their possessor (Beenfeld, 2013; Leahey, 2014). Accepting this, Kant concluded that psychology could not be a science because its object of study could not be empirically observed. However, it has proved possible to infer, even when we cannot see, mental processes. The technique is called inference to the best explanation, and interestingly it’s based more on law than science.

如果你看过侦探剧，你一定见过推理是如何得出最佳解释的。除非闭路电视或手机拍下了犯罪过程，或者有目击证人，否则犯罪事件必须根据无声证据推断出来。侦探和律师会仔细研究犯罪线索，收集证词、法医证据等等，拼凑出一个犯罪故事。关于谁、什么、何时、何地以及为什么，可能存在不止一种解释，但通常其中一种解释会成为最佳解释，案件会被提交审理，尽管辩方可能会提出另一种犯罪理论。

If you have watched detective shows you have seen inference to the best explanation at work. Unless CCTV or someone’s phone has caught a crime on video, or there are eyewitnesses, the events of a crime must be inferred from mute evidence. Detectives and lawyers go over clues to the crime, collect statements, forensic evidence, and so on to put together a story of the crime. There may be more than one possible explanation of who, what, when, where, and why, but usually one of them emerges as the best explanation and the case is brought to trial, though the defense may offer an alternative theory of the crime.

我们在认知心理学中也做着同样的事情，而且还具备实验的优势。与侦探或历史学家不同，科学家可以创造新的事实来阐明隐藏的过程。一个关于工作记忆的经典实验（Sternberg，1966）就是一个很好的例子。

We do the same thing in cognitive psychology, with the added advantage of running experiments. Unlike a detective or a historian, a scientist can create new facts to illuminate hidden process. A classic experiment on working memory (Sternberg, 1966) provides an elegant example.

当时已确定工作记忆的容量约为七项（Miller，1956）。Sternberg 想了解人们如何搜索存储在工作记忆中的一组项目。在一系列试验中，Sternberg 向他的参与者（当时称为“受试者”）展示了屏幕上一秒多一点的 1 到 6 个数字的列表。然后，参与者看到一个探测数字，并被问及该探测数字是否在列表中；一半时间答案是肯定的，一半时间答案不是。Sternberg 测量了延迟时间，即响应时间：参与者花了多长时间回答这个探测问题？这项任务非常简单，几乎没有出现任何错误。

It had already been established that the capacity of working memory was about seven items (Miller, 1956). Sternberg wanted to find out how people search a set of items stored in working memory. In a series of trials, Sternberg presented his participants (“subjects” in those days) a list of one to six numbers on a screen for just over a second. Then, participants saw a probe digit and were asked whether the probe digit was on the list; half the time it was and half the time it wasn’t. Sternberg measured latency, or response time: How long did it take participants to answer the probe question? The task is very easy and was carried out virtually without error.

斯特恩伯格发现，在负面探测（不在列表中）中，列表大小决定了潜伏期的长度：列表越长，回答“否”所需的时间就越长。这表明人们会一次搜索一个存储的条目列表，即串行搜索。在负面试验中，这意味着要搜索记忆中的每个条目，因此必须搜索整个列表才能做出反应。更令人惊讶的是，同样的模式也出现在正向探测试验。列表越长，参与者回答“是”所需的时间越长。人们可能认为探测在列表中的位置会决定反应时间，因此如果探测出现在列表中的第二个位置，它会比出现在第四个位置时更快地被检测到。这本来是一个串行的自我终止搜索，在找到探测时结束，无需继续寻找已经找到的内容。然而，即使检测到了探测，参与者似乎仍然搜索了整个列表。正向和负向搜索都是详尽的串行搜索。

Sternberg found that on negative probes (it was not in the list) list size determined the length of latency: the longer the list, the longer it took to respond “No.” This indicates that people search a stored list of items one at a time, i.e., serially. On a negative trial, this meant searching every slot in memory for the probe, so that the whole list had to be searched before responding. More surprising was that the same pattern emerged on the positive probe trials. The longer the list, the longer it took participants to respond “Yes.” One might have expected that a probe’s position in the list would have determined reaction time, so that if the probe had occurred in the list at position two it would be detected more quickly than if it were at position four. This would have been a serial self-terminating search, ending when the probe is found, there being no need to continue looking for what had already been found. Instead, it appeared that even when a probe was detected, participants nevertheless searched the whole list. Both positive and negative searches were exhaustive serial searches.

因此，就像在刑事调查中一样，心理侦探斯滕伯格无法直接观察参与者的搜索过程，但他可以通过可测量的反应时间推断出来。通过进行多个实验（我这里只描述了一个），他排除了相互竞争的理论，并最终找到了最佳解释。

Thus, as in a criminal investigation, Sternberg, the psychological detective, could not directly observe his participants’ search processes, but he could infer them from measurable reaction times. By running multiple experiments (I have described here just one), he eliminated rival theories and zeroed in on the best explanation.

比较和对比系统

Comparing and Contrasting the Systems

效仿卡尼曼，我们可以通过考察两个典型案例来阐明每个系统的关键特征。想象一下，首先，你看到一张愤怒的脸出现在你面前。其次，快！在心算中计算 27 × 13 = 。

Following Kahneman, we can illustrate the key features of each system by examining two paradigmatic examples of each at work. Imagine, first, seeing an angry face thrust before you. Second, quick! In your head multiply 27 × 13 = .

系统 1 是无意识的，其运行快速且自动。当你看到一张愤怒的脸时，你立刻就知道此人生气了。你不会有意识地从线索推断出愤怒，而是体验到愤怒。你的大脑确实会使用线索来推断他的愤怒，但这些计算是深度无意识的。该过程是如此自动，它很可能是系统 1 中一个独立运行的模块化子系统的示例。与此形成鲜明对比的是，简单的算术问题是有意识地完成的，需要集中的脑力劳动，并且几乎任何其他活动都可能打断它。卡尼曼指出，如果你和朋友一起散步，并被要求计算 27 × 13，你几乎肯定会停下脚步去做算术！系统 2 过程需要深思熟虑才能遵循学到的算术规则。另一方面，最重要的（但不是唯一）系统 1 过程是联想记忆，它执行联想定律。从心理学发展初期开始，心理学家就区分了两种注意力：非自主注意力（即外部刺激吸引注意力）和自主注意力（即主动努力去关注重要刺激，例如聚会上的对话）。这两种注意力分别对应系统1和系统2。

System 1 is unconscious, and its operations are fast and automatic. When you see an angry face, you know immediately that the person is angry. You do not consciously work out from cues to anger, you experience it. Your brain does use cues to infer his anger, but these computations are deeply non-conscious. The process is so automatic, it’s likely an example of a modular sub-system of System 1 that operates on its own. In strong contrast, the simple arithmetic problem is done consciously, requiring concentrated mental effort, and it can be disrupted by almost any other activity. Kahneman points out that if you were strolling with a friend, and were asked to multiply 27 × 13, you would almost certainly stop walking to do the math! System 2 processes require deliberation in order to follow the learned rules of arithmetic. The most important (though not the only) System 1 process, on the other hand, is associative memory, which executes the laws of association. From the earliest days of psychology, psychologists distinguished two forms of attention, involuntary, when an external stimulus captures your attention, and voluntary, when you make an active effort to attend to an important stimulus, such as a conversation at a party. These types of attention map onto System 1 and System 2, respectively.

笛卡尔将他的意识，即他的系统2，视为他的本质自我，而我们往往也认同他的直觉。我们承担和分配道德责任，很大程度上是基于有意识的行为意图。因此，一级谋杀需要“预谋恶意”，而如果一个人能够合理地辩称自己在情绪压力下“突然崩溃”，他们可能会被判过失杀人罪，甚至可能被无罪释放。至少所有哺乳动物都具备感知威胁的能力，而且我们通常从动物身上继承了系统1。祖先。系统 2 中人类独有的心理过程很大程度上是由我们的文化通过模仿和正式学习而建立的 (Boyer, 2018; Tomasello, 2019)。

Descartes regarded his consciousness, his System 2, as his essential self, and we tend to share his intuition. We take and assign moral responsibility largely on the basis of conscious intention to act. Thus, first-degree murder requires “malice aforethought,” while if a person can plausibly argue that they “just snapped” under emotional strain, they may be found guilty of manslaughter or even acquitted. The ability to detect threat displays is found in at least all mammals, and in general we inherit System 1 from our animal ancestors. The distinctively human mental processes of System 2 are largely installed by our cultures, through imitation and formal learning (Boyer, 2018; Tomasello, 2019).

卡尼曼将系统 1 的工作概括为不断构建世界模型来指导我们的行为。我们无法选择阻止模型的构建和呈现——即使我们被告知视觉错觉是错觉，它们也不会消失，而表演出色的魔术也会愚弄那些知道如何表演魔术的魔术师；对魔术的最高赞美是“我甚至欺骗了我自己！”最重要的是，我们心理生活的世界是由系统 1 构建的世界。卡尼曼将此事实称为 WYSIATI：你所看到的就是这里的一切。 在 本书 中 ，我们将了解卡尼曼在提出预测处理假设时低估的一个WYSIATI 方面。我们生活的世界不是世界的模型，而是系统 1 预测的不久的将来的世界的模型。这听起来令人震惊，但正如卡尼曼强调的那样，大象非常擅长它的工作。它的预测和预期反应几乎总是正确的。骑象人可以放心地让大象独自一人，愉快地在它的背上放松。只有当预测失败时，系统1才需要唤醒系统2，而解决新问题则需要有意识的思考。它们之间的关系具有高度的适应性，因为从生物学角度来看，系统1的运行成本很高，正如我们在算术问题中看到的那样。

Kahneman summarizes the job of System 1 as continuously building a model of the world for guiding our behavior. We can’t choose to stop the model from being built and presented—optical illusions don’t go away even when we are taught that they are illusions, and good magic tricks performed well nevertheless fool magicians who know how they are done; the highest praise for a magic trick is, “I even fool myself!” Most importantly, the world we psychologically live in is the world constructed by System 1. Kahneman dubs this fact WYSIATI: W hat Y ou S ee I s A ll T here I s. In this book, we will meet an aspect of WYSIATI that Kahneman under-appreciated when he suggested the predictive processing hypothesis. The world we live in is not a model of the world, it is a model of the world as System 1 predicts it to be in the immediate future. This sounds alarming, but, as Kahneman stresses, the elephant is very good at its job. Its predictions and anticipated responses are almost always right. The rider can simply leave the elephant alone, happily relaxing on its back. System 1 need wake System 2 only when a prediction fails, and conscious thought is needed to solve a new problem. Their relationship is highly adaptive because System 1 is biologically expensive to run, as we saw with the arithmetic problem.

关于隐喻和模型的注释

Notes on Metaphors and Models

到目前为止，我们已经在本书中讨论了许多关于心灵的隐喻，例如柏拉图的战车、双系统地图以及海特的大象和骑象人。你可能还听过其他一些隐喻，例如弗洛伊德的马（本我）和骑象人（自我），或者心灵是冰山一角，大部分都隐藏在水面之下，当然，还有如今无处不在的计算机。在其他心理学课程、书籍或文章中，你也会遇到“模型”。它通常被用来指适用于少数现象的小规模理论，例如斯滕伯格的短期记忆序列穷举搜索模型。然而，我们需要稍微谈谈隐喻和模型，因为前者可能会悄悄地误导人，而后者在科学哲学中的用法比心理学家给出的更为精确。

We have discussed a number of metaphors for mind in this book so far, Plato’s chariot, the two-system map, Haidt’s elephant and rider. You may have heard others, such as Freud’s horse (id) and rider (ego) or mind as iceberg with most below the water’s surface, and, of course, the now-ubiquitous mind as a computer. In other psychology courses, books, or articles, you have met “models,” too. It’s usually deployed to mean a small-scale theory applying to a small number of phenomena, such as Sternberg’s serial exhaustive search model of short-term memory. However, we need to say a little about metaphors and models because the former can be quietly misleading, and the latter has a more precise use in philosophy of science than psychologists give it.

我们总是用比喻来表达难以捉摸的想法和感受，但它们可能难以捉摸，有时意图比表面更多，有时更少。“我的爱就像一朵红红的玫瑰”，听起来不错，但作者是想让我们思考它的刺吗？还是过敏？“我能否将你比作夏日？”炎热潮湿，午后还可能有雷雨？在科学等不那么浪漫的语境中，比喻是引导人们思考新奇陌生事物的有效方式，如果直接呈现，这些事物可能会令人望而生畏。因此，老师会将新事物与旧事物进行比较，例如将心智的双系统观比作大象和它的骑手。Haidt 的比喻也利用了你可能已经知道的比喻，例如弗洛伊德的马和骑手。海特故意用大象代替马，以强调大象的体型——系统 1 比系统 2 运行更多的流程——以及与马相比，大象的体型使其更难驾驭的事实。

We use metaphors all the time to help us express elusive ideas and feelings, but they can be slippery, intending sometimes more, sometimes less, than what’s on the surface. “My love is like a red, red rose,” sounds nice, but did the author want us to think about the thorns? Or allergies? “Shall I compare thee to a summer’s day?” Hot and humid with a chance of afternoon thunderstorms? In less romantic contexts such as science, metaphors are useful ways of inviting someone to consider something new and unfamiliar that, presented directly, might seem daunting. So, a teacher will compare a new thing to an old one, such as the two-system view of the mind to an elephant and its rider. Haidt’s metaphor also takes advantage of metaphors you may already know, such as Freud’s horse and rider. Haidt deliberately replaced the horse with the elephant to emphasize the elephant’s size—System 1 runs many more processes than System 2—and the fact that, compared to horses, their bulk makes elephants harder to steer.

然而，在任何隐喻中，新事物和它的隐喻之间都会有相似之处和不同之处；有些是有意为之，有些则不是，但就像玫瑰的刺一样，其他的可能以模棱两可的状态潜伏在背景中。Haidt 有意表现大小相似，但他肯定不是在暗示系统 2 有一个可卷曲的鼻子。但是大象的大耳朵呢？Haidt 的意思是系统 1 对环境特别警觉，能够意识到小耳朵的系统 2 所没有的东西吗？事实上，这个问题是注意力心理学中的一个严肃问题，但目前尚不清楚 Haidt 是否对此表明了立场。最后，隐喻可能会暗示提出者从未想过的某些思路。在印度，大象通常不仅载着骑手，还载着乘客，比如告诉骑手去哪里的王公。添加大君会改变这个比喻，从而为许多人从认知科学中得出的令人不安的推论提供新的解释，即我们没有自由意志。

In any metaphor, however, there will be similarities and differences between the new thing and its metaphor; some will be intended and some not, but as with the rose’s thorns, others may lurk in the background in an ambiguous state. Haidt intended the similarity of size, but he surely did not mean to imply that System 2 has a prehensile trunk. But what about the elephant’s big ears? Did Haidt mean to imply that System 1 is especially alert to the environment, being aware of things the small-eared System 2 isn’t? This question is, in fact, a serious one in psychology of attention, but it’s not clear if Haidt is taking a position on it. Finally, metaphors may suggest certain lines of thought never meant by their proposer. In India, elephants often carried not just a rider, but a passenger such as a maharaja who tells the rider where to go. Adding the maharaja changes the metaphor in a way that throws new light on a disturbing implication that many people draw from cognitive science, that we have no free will.

模型是一种隐喻，因为它部分地表征了某个对象。但在科学哲学中，它们还有更严谨的含义，超越了心理学家将“模型”用作小层次理论的范畴（Suppe，1989）。模型是对世界将会是什么样子的一种表征，前提是该理论所假设的对象和力恰好是世界上起作用的。这种用法始于伽利略从物理学中摒弃次要的感觉属性。由于它们不影响物体的因果相互作用，因此运动的数学动力学可以安全地忽略它们。但伽利略也忽略了其他确实会影响运动的因素，例如摩擦力，包括空气阻力。在高中物理课上，你可能学过如何解斜面问题，其中摩擦力以及物体的形状和物质与你对其加速度的计算无关。严格来说，科学的预测和解释是关于理论模型的理想化世界，而只是间接地或隐喻地关于现实世界。

Models are metaphors in that they partly represent an object. But in philosophy of science, they also have a more rigorous meaning that also goes beyond psychologists’ use of “model” as small-level theory (Suppe, 1989). A model is a representation of what the world would be like if only and exactly the objects and forces posited by the theory were the ones operative in the world. This usage began to emerge with Galileo’s dismissal of the secondary sense properties from physics. Because they did not affect objects’ causal interactions, they could be safely ignored by the mathematical dynamics of motion. But Galileo ignored other factors, too, that do affect motion, such as friction, including air resistance. In high school physics, you probably learned how to do inclined plane problems, where friction and the shape and matter of an object don’t matter to your calculations of its rate of acceleration. Scientific predictions and explanations are, strictly speaking, about a theoretical model’s idealized world and only indirectly, or metaphorically, about the real world.

心理学和其他社会科学领域也是如此。斯滕伯格的模型（隐含地）是关于一位身体状况良好、动机十足的参与者专注于记忆任务。有些人可能感到无聊、宿醉或还没喝咖啡，但这些动机因素并不在斯滕伯格的模型范围内。这就是为什么需要进行统计分析，并对许多参与者进行多次试验和设置条件，以便让这些特质因素相互抵消。事实上，其他科学家可以使用像斯滕伯格这样的经过充分测试的程序来研究药物对行为的影响或评估脑损伤等问题，因为它们提供了正常行为的基线。

The same is true in psychology and other social sciences. Sternberg’s model was (implicitly) about a perfectly well-motivated participant in good physical shape paying sole attention to the memory task. Some may have been bored, hung-over, or hadn’t had their coffee yet, but such motivational factors were outside Sternberg’s model. This is one reason for statistical analysis and running many trials and conditions with many participants, to allow such idiosyncratic factors to cancel each other out. In fact, well-tested procedures such as Sternberg’s can be used by other scientists to study things, such as the effects of drugs on behavior, or to assess brain damage, because they provide baselines of normal behavior.

然而，认知科学关于模型及其应用的标准假设也带来了一些挑战。例如，最近人们普遍认为，心理学家倾向于认为文化对许多事情的影响并不大。人类行为。认知科学和社会科学研究中的绝大多数受试者来自北美，其中也混杂着西欧人。然而，所有这些受试者都是WEIRD型（Henrich、Heine & Norenzayan，2010）。也就是说，他们都成长并生活在西方、受教育程度高、工业化程度高、富裕且民主的国家。许多重复性良好的实验都针对WEIRD型受试者群体进行，但或许正是因为他们的社会背景相似，也就是说，这是一个无意的控制变量，导致我们对人性的看法可能存在狭隘性。

However, certain challenges arise from cognitive science’s standard assumptions about models and their use. For example, it has lately been accepted that psychological scientists have tended to assume that culture was not an important influence on much of human behavior. The vast, vast majority of people studied in cognitive science and social science are drawn from North America, with a mixture of western Europeans. However, all these participants are WEIRD (Henrich, Heine, & Norenzayan, 2010). That is, they all grew up and live in countries that are Western, Educated, Industrialized, Rich, and Democratic. Many well-replicated experiments have been done with the WEIRD participant pool, but it just may be that because their social background is similar, i.e., it has been an inadvertent control variable, our view of human nature may be parochial.

最后，工程师，例如认知科学应用分支——人工智能领域的工程师，不能忽视理论模型中忽略的变量。埃及工程师沿着堤道搬运方形石块建造金字塔，新石器时代的工程师将巨大的砂岩搬运到巨石阵，罗马工程师搬运石块修建道路和渡槽，巴黎圣母院的建造者将石块运到现场进行雕刻，他们不可能仅仅假设摩擦力就消除。他们必须克服它。人工智能程序必须成功运行才能在真实的高速公路上安全行驶，就像自动驾驶汽车项目的目标一样。认知科学与实用人工智能认知工程之间的差异，在优雅但受限的科学模型与现实世界的混乱之间，形成了创造性的张力。

Finally, engineers, such as the engineers in cognitive science’s applied branch, AI, cannot ignore the variables ignored in theoretical models. The Egyptian engineers who hauled square stones around a causeway to build the pyramids, the Neolithic engineers who hauled the great sarsen stones to Stonehenge, the Roman engineers who moved stones to build their roads and aqueducts, and the cathedral builders of Notre Dame bringing stone on-site for sculpting could not have just assumed friction away. They had to overcome it. Programs in artificial intelligence must be successful in order to safely drive on real highways, as self-driving car projects aim to do. The difference between cognitive science and practical AI cognitive engineering makes a creative tension between the elegant but restrictive models of science and the messiness of the real world.

我们现在的情况

Where We Are Now

一些科学家反对双系统模型（Melnikoff & Bargh, 2018a；Pennycook 等人，2018 对此进行了辩护；Melnikoff & Bargh, 2018b 对此进行了回应），认为它不足以支撑认知科学，这当然是错误的。它就像一个比喻，一张探索心智的松散地图，对我们大有裨益。最终，它会被更精确的 GPS 所取代。

Some scientists reject the two–system model (Melnikoff & Bargh, 2018a; defended by Pennycook et al., 2018; reply by Melnikoff & Bargh, 2018b) as an insufficient theory for cognitive science, which of course it is. It’s a metaphor, a loose map for exploring the mind that will serve us well. Eventually it can be replaced by a more precise GPS.

外卖

Takeaway

我们有时会想象问一只蜈蚣一个糟糕的问题：“一百条腿怎么走路？”这对蜈蚣来说简直是灾难，因为它一想到要走路，就感到困惑，根本走不了。走路最好留给大象去处理，留给系统1的无意识处理。

We sometimes imagine asking a centipede a disastrous question, “How do you walk with a hundred legs?” It’s a disaster for the centipede because as soon as it thinks about walking, it gets confused and can’t walk. Walking is best left to the elephant, to the unconscious processing of System 1.

但如果你像我一样失去了行走的能力，你必须重新思考如何才能再次行走。物理治疗就是学习如何有意识地去做你一生中无意识地做的事情。电影和电视对物理治疗的描述往往具有误导性或错误性。一个常见的桥段是，警察或士兵受伤后，在双杠上重新学习行走，放弃希望，而朋友、爱人或严厉的老板会大喊：“你能行！你必须要有决心！” 在一个感人的时刻，脚步又回来了，这是意志的胜利。事实并非如此，而是一路走来的小步。今天两步，明天四步，下周十八步，直到很久以后（对我来说是三个月），你才能正常行走而不会摔倒。一路走来，你必须时刻警惕自己和周围环境的细微变化，这些变化你以前从未考虑过：我能抓住哪里支撑？我能走多少步才会倒下？

But if you lose the ability to walk, as I did, you must think about it to do it again. Physical therapy is about learning to do consciously what you spent your life doing unconsciously. Movie and TV depiction of physical therapy is often misleading or wrong. A common trope is a cop or soldier learning to walk again after being wounded, on parallel bars, giving up hope while a friend, lover, or stern boss-type yells “you can do it! You have to want it!” And in a touching moment, steps return, a triumph of the will. It’s not like that, but is baby steps all the way. Two today, four tomorrow, 18 next week, until after a long time (for me, three months) you can walk normally without falling over. Along the way, you must be alert to nuances of yourself and your surroundings you never computed before: Where can I grab for support? How many steps can I take before I might collapse?

另一个误解发生在坐轮椅的角色突然恢复行走的场景中。例如，在《骗徒大亨》中，史蒂夫·马丁假装自己是一位坐轮椅的老兵，（出于剧情原因）必须站起来行走。就像其他电影中的描述一样，他直接从轮椅上跳起来，然后一步一步地、僵硬地、机械地行走。事实上，行走是容易的部分，站起来才是难的。在正常情况下，我们会不假思索地在椅子上向前倾，双手轻轻地扶住椅子的扶手、沙发的座位或床沿，以此来准备站立。这样做可以让我们的重心集中在双脚上，这样双腿就只需要向上推即可。然而，如果我们身体虚弱或感到疼痛，我们往往会忽略这些动作，即使它们比健康时更有必要，因为我们需要将身体的重心恰好而不是大致地放在双脚上。物理治疗师会教我们一句格言：“鼻子比脚趾重！”站起来之前，你需要在椅子上向前倾，直到鼻子和脚趾在地面上对齐，然后再尝试站起来。一开始，鼻子和脚趾的距离不能差一毫米，否则你就站不起来。一旦站起来，膝盖可以锁紧，就可以走路了。虽然仍然很难，但还是有可能的。

Another misconception occurs in scenes where a wheelchair-bound character suddenly walks again. For example, in Dirty Rotten Scoundrels, Steve Martin is faking being a wheelchair-bound veteran who (for plot reasons) has to get up and walk. As in other cinema depictions, he bolts straight out of his wheelchair and then walks badly, and robotically, rigid step by rigid step. In fact, walking is the easy part; getting up is the hard part. Without thinking about it, in our normal state we prepare to get standing by leaning forward in our chair and bracing, if lightly, our hands on the arms of the chair or the seat of a sofa or the edge of our bed. Doing so gets our center of gravity over our feet, so our legs just have to push up. However, if we are weak and/or in pain, we tend to neglect these moves, even though they are more necessary than when we were healthy because we need our weight to be exactly rather than approximately over our feet. Physical therapists teach a motto: “Nose over toes!” Before standing up, you lean forward in your chair until you visually align your nose with your toes on the floor, and then work on standing up. Initially, your nose can’t be short of your toes by even a millimeter, or you’ll fail to stand. Once up and your knees can lock, walking is possible. Still hard, but possible.

通过物理治疗，我对我的大象产生了极大的敬意，它在背后努力工作，让我能够继续前进，这样我的意识就可以做其他事情了。

Going through physical therapy gave me enormous respect for my elephant, working hard in the background to keep me going so my conscious mind could do other things.

阅读材料

Readings

关于人工智能的历史，可以参考M. Woolridge的《人工智能简史：它是什么，我们现在在哪里，以及我们要去哪里》（伦敦：Macmillan出版社，2021年），该书强调了失败，旨在展示科学家的思维方式。J. Shane的《你看起来像个东西，而我爱你：人工智能是如何运作的，以及它为何让世界变得更奇怪》（波士顿：Little, Brown出版社，2019年）则以幽默但严肃的视角审视了人工智能。Shane也关注那些有趣的失败，这些失败揭示了思维的复杂性，以及人工智能的难度。她运营着一个人工智能幽默博客：https://www.aiweirdness.com。

A history of AI is M. Woolridge, A Brief History of Artificial Intelligence: What It Is, Where We Are, and Where We Are Going (London: Macmillan 2021), which emphasizes failures and aims to show how scientists think. A funny but ultimately serious look at AI is provided by J. Shane, You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place (Boston: Little, Brown, 2019). Shane, too, focuses on failures—amusing ones—that reveal how complicated the mind is and thus how hard AI is. She runs an AI humor blog: https://www.aiweirdness.com.

M. Lewis 的《毁灭计划：改变我们思想的友谊》（纽约：诺顿，2016 年）是丹尼尔·卡尼曼和他的长期学术伙伴阿莫斯·特沃斯基共同撰写的传记，作者是著名作家（《点球成金》、《大空头》）。

M. Lewis, The Undoing Project: A Friendship That Changed Our Minds (New York: Norton, 2016) is a joint biography of Daniel Kahnemann and his long-time academic partner, Amos Tversky, by a well-known writer (Moneyball, The Big Short).

电影

Movies

《科学》杂志邀请了一些人工智能科学家挑选出最佳的人工智能电影，并对其科学准确性进行评分。以下是名单（现在看来有些过时了）：https://www.sciencemag.org/news/2015/07/which-movies-get-artificial-intelligence-right。

The journal Science asked some AI scientists to pick the best AI films and rate them for scientific accuracy. Here’s the list (a bit dated now): https://www.sciencemag.org/news/2015/07/which-movies-get-artificial-intelligence-right.

我选《2001太空漫游》最好，因为HAL就是在我所在的伊利诺伊大学研究生院发明的。9 作为一个科幻迷，我读过艾萨克·阿西莫夫（《机器人总动员》）和阿瑟·C·克拉克（《2001太空漫游》）的^作品。我从小就以为《2001太空漫游》里的世界至少会像库布里克电影里展现的一样先进。正如库布里克在《卡萨布兰卡》里谈到那里缺水时所说，“我被误导了。” ¹⁰

I’ll go with 2001 as the best, if only because HAL was created at my graduate school, University of Illinois.⁹ A sci-fi fan reading Isaac (I Robot) Asimov and Arthur C. (2001) Clarke, I grew up thinking that the world in 2001 would be at least as advanced into space as shown in Kubrick’s film. As Rick says in Casablanca about the lack of waters there, “I was misinformed.”¹⁰

魔法

Magic

本章没有特别的魔术，所以我推荐一部关于20世纪伟大魔术师之一的电影：《欺骗之道：瑞奇·杰伊的秘密与导师》（2012）；以及《惊天​​魔盗团》（2013）。这些魔术都是真实存在的。

There’s no special magic for this chapter, so I’ll recommend a movie about one of the great 20th-century magicians: Deceptive Practice: The Mysteries and Mentors of Ricky Jay (2012); and Now You See Me (2013). The tricks are real.

建筑 I：符号系统

Architecture I: Symbol-Systems

思想的语言

The Language of Thought

笛卡尔说，语言是内在和私人灵魂的外在可见的符号。图灵摒弃了灵魂，但在他的模仿游戏中，他保留了语言。用作精神的标志。本质上，对于图灵及其随后的符号系统架构而言，思考是内心的对话，而智力是思考的语言（Fodor，1975）。

Descartes said that language was the outward and visible sign of the inward and private soul. Turing dispensed with the soul, but, in his imitation game, he retained language use as the mark of the mental. In essence, for Turing and the symbol system architecture that followed, thinking was inner talking, and intelligence is the language of thought (Fodor, 1975).

让我们以思维语言在工作中的作用为例，并用它来形式化符号系统假说。一个夏日的傍晚，你和一位朋友聊天，对方建议你如果明天天气好的话，就去公园烧烤。第二天，你醒来，图 3.1显示了你大脑中发生的事情，这符合认知的符号系统假说，并观察到与表 2.1 的相似之处。

Let’s take an example of the language of thought at work and use it to formalize the symbol system hypothesis. You’ve spoken to a friend one summer evening, and it’s suggested that if the weather stays pleasant tomorrow, you’ll go barbequing in the park. Next day, you wake up, and Figure 3.1 shows what happens in your mind, according to the symbol system hypothesis of cognition, observing the similarity to Table 2.1.

处理计算易处理性：模块化

Dealing with Computational Tractability: Modularity

福多尔 (1983) 提出了解决这两个问题的方法，即心智模块性，这一理论影响了感知神经科学和进化心理学（图 3.2）。

Fodor (1983) proposed a solution to both problems, modularity of mind, that has influenced perceptual neuroscience and evolutionary psychology (Figure 3.2).

福多尔的模块化理论适用于信息处理的早期阶段，类似于亚里士多德的特殊感官。福多尔的理论认为，大脑中专门的区域只处理呈现给每个感觉器官的信息。实际上，人脑中并非像个人电脑那样只有一个计算机芯片，而是有多个芯片，每个芯片负责处理每种感觉，再加上一个通用的中央处理器（CPU）来整合各种感觉流。模块化提高了处理速度，因为每个感觉模块只负责一种数据，并且每个模块都是封装的，这意味着其处理不会影响其他模块的处理，也不会受到其他模块处理的影响，这些模块也不受中央处理器的控制。

Fodorian modularity applied to the earliest stages of information processing and resembles Aristotle’s special senses. Fodor’s theory held that dedicated tracts of the brain process only the information presented to each sense organ. In effect, instead of there being a single computer chip in the human head as in a PC, there were several, one for each sense, plus a general-purpose Central Processing Unit that brought the sensory streams together. Modularity gained processing speed because each sensory module was responsible for only one kind of data, and each module was encapsulated, meaning that its processing did not affect and was not affected by the processing in other modules, nor were the modules under control of the central processor.

为了说明他的意思，福多尔指出了感官错觉，例如艾宾浩斯/铁钦纳视觉错觉（图 3.3）。

To illustrate what he meant, Fodor pointed to sensory illusions, such as the Ebbinghaus/Titchener visual illusion (Figure 3.3).

在这个错觉中，两个显示器的中心圆圈大小相同，但由于与周围较大或较小的圆圈形成对比，看起来有所不同。对福多尔来说，重要的是，即使你被告知错觉的原理并测量了圆圈，你也无法消除错觉。它是与生俱来的、不可避免的，它提供了封装视觉模块存在的证据。

In this illusion, the central circles in both displays are the same size but appear to differ due to the contrast with the larger or smaller circles surrounding them. What is important to Fodor was that even if you are told how the illusion works and you measure the circles, you can’t make the illusion go away. It’s innate and inescapable, providing evidence of the existence of an encapsulated visual module.

福多尔的观点巧妙地缓解了实时性问题。它也朝着将心智视为并行进行信息处理的理念迈出了一步，引领我们走向下一个认知架构——联结主义。

Fodor’s notion was a clever way of ameliorating the real-time problem. It was also a step toward thinking of the mind as performing information processing in parallel, taking us part of the way to the next architecture of cognition, connectionism.

架构 II：联结主义和预测处理

Architecture II: Connectionism and Predictive Processing

在20世纪50年代认知科学的萌芽阶段，人们提出了两种机器智能的形式。一种是符号系统方法。另一种是通过所谓的“神经元网络”来探索的。其最重要的倡导者是心理学家兼工程师弗兰克·罗森布拉特（Frank Rosenblatt，1928-1971）。他构建了一种名为感知器（Perceptron）的电子神经网络（Rosenblatt，1958），能够学习识别字母等简单模式。直到美国国防高级研究计划局（DARPA，该局也创建了互联网，即阿帕网）认定美国未来的军事命运取决于信息处理技术的进步，并开始向新兴的人工智能领域投入资金，这两种方法之间的竞争才变得激烈起来。符号系统观点的支持者认为感知器无法处理某些基本的逻辑功能，因此注定会失败（Minsky & Papert，1969）。尽管事实证明他们错了，但对感知器的支持逐渐消失，直到 20 世纪 80 年代联结主义/神经网络复兴（Olazaran，1996）才重新出现，从而带来了当今的深度学习和大数据时代。

At the dawn of cognitive science in the 1950s, two forms of machine intelligence were proposed. One was the symbol system approach. The other was sought through what were called “neuron nets.” Its foremost advocate was psychologist and engineer, Frank Rosenblatt (1928–1971). He built an electronic neural network called the Perceptron (Rosenblatt, 1958) capable of learning to recognize simple patterns such as letters. Competition between the two approaches did not become serious until DARPA (Defense Advanced Research Project Agency, which also started the Internet as ARPANET) decided that America’s future military fortunes depended on advances in information-processing technology and began to throw money at the budding field of AI. Advocates of the symbol-system view argued that Perceptrons could not handle certain basic logical functions and were thus doomed to failure (Minsky & Papert, 1969). Although it turned out they were wrong, support for Perceptrons withered away, only to return with the connectionist/neural net revival of the 1980s (Olazaran, 1996), leading to today’s eras of deep learning and Big Data.

基本主张：心理过程是联想强度的计算

Essential Claim: Mental Processes Are Calculations of Association Strengths

联结主义者的核心论点是，符号系统体系结构对 Marr 的第三个逆向工程层次——实现——关注太少。在常规工程中，人们最终会构建设备来执行所提出的算法。然而，在认知科学中，我们必须能够证明我们的算法可以在已经存在的物理系统——人脑中实现。显然，人脑太慢，无法实现思维的串行处理，但它比任何计算机都好。此外，它还具有串行计算机所不具备的优点，例如优雅降级。通常，如果标准数字计算机中一个部件出现问题，整个系统都会崩溃。然而，人脑可以承受一定程度的损伤并继续运作，尽管功能会减弱或受限。

Connectionists’ central argument was that the symbol-system architecture had paid too little attention to Marr’s third level of reverse engineering, implementation. In regular engineering, one ends by building the device to carry out proposed algorithms. However, in cognitive science we must be able to show that our algorithms can be implemented in an already existing physical system, the human brain. It is obvious that the human brain is too slow to implement serial processing of its thoughts, yet it works better than any computer. Moreover, it has virtues that serial computers do not, such as graceful degradation. Generally, if one thing goes wrong in a standard digital computer, the whole system fails. However, the human brain can take degrees of damage and still operate, albeit in weakened or limited forms.

有趣的是，联结主义也回顾了一种更古老的智力模型——联想主义，如今这种模型可以用神经中心主义的语言来表达。智力是由网络中自适应变化的连接构成的，这些连接由简单的处理单元组成，这些单元是生物神经元的数学模型。

Interestingly, connectionism also looked backward to an older model of intelligence, associationism, expressed now in neurocentric language. Intelligence consists of adaptively changing connections in a network consisting of simple processing units that are mathematical models of biological neurons.

图 3.4展示了一个利用声纳区分水雷和水下岩石的简单网络，图 3.5展示了联结网络中一个典型的“神经元”。在神经系统中，突触连接有两种类型：兴奋性连接和抑制性连接。兴奋性连接是指突触前神经元（图中未显示）的放电倾向于引起目标神经元的放电；抑制性连接是指突触前神经元的放电倾向于阻止目标神经元的放电。图中，兴奋性连接用箭头表示，抑制性连接用小圆圈表示。每个连接都关联一个权重，其值从 0（无影响）到 1（最大影响）；因此，兴奋性连接的值从 0 到 1，抑制性连接的值从 0 到 -1。处理单元（或“神经元”）只需将其输入值相加，并计算一个指定的函数。在图 3.5中，我输入了一个简单的阈值阶跃函数，就像实际神经元通常计算的那样。如果总输入等于或大于设定的阈值，该单元就会触发，向其突触后神经元（未显示在目标神经元上方）发送兴奋性和/或抑制性冲动。如果未达到阈值，该单元将保持沉默。

Figure 3.4 shows a simple network for using sonar to distinguish mines from underwater rocks, and Figure 3.5 shows a typical “neuron” in a connectionist network. In the nervous system, synaptic connections are of two types, excitatory, in which the firing of the presynaptic neuron (not shown in the Figure) tends to cause the target neuron to fire, and inhibitory, in which firing of the presynaptic neuron tends to prevent the target neuron from firing. In the Figure, excitatory connections are indicated by arrows, inhibitory connections by small circles. Each connection has a weight associated with it, whose value runs from 0 (no effect) to 1 (maximum effect); thus, excitatory values run from 0 to 1 and inhibitory from 0 to -1. The processing unit, or “neuron,” simply sums the values of its inputs and computes an assigned function. In Figure 3.5, I have put in a simple threshold step function such as actual neurons typically compute. If the total input is equal to or greater than the set threshold, the unit fires, sending excitatory and/or inhibitory impulses to its postsynaptic neurons (not shown above the target neuron). If the threshold is not met, the unit remains silent.

在图 3.4中，数量不定的单元被组装成一个典型的多层前馈网络。该网络由三层组成：输入层，由神经元组成，这些神经元的响应由每个神经元接收到的感觉刺激强度决定；隐藏层，将输入单元连接到输出单元；输出层，其触发反映了网络的总体活动。虽然为了清晰起见没有显示出来，但每个输入单元都与每个隐藏单元都有连接，每个隐藏单元都与每个输出单元都有连接。它被称为前馈网络，因为连接的方向仅从输入到输出。没有连接同一层级的单元之间缺乏连接，后级单元与前级单元之间缺乏刺激连接。每个单元都完全相同，没有中央处理的空间。

In Figure 3.4, an unspecified number of units has been assembled into a typical multilevel feed-forward network. This one is made up of three layers, an input layer, consisting of neurons whose responses are determined by the strength of sensory stimulation each one receives, a hidden layer that connects input units to output units, and an output layer whose firing reflects the aggregate activity of the network. Although it’s not shown for clarity, every input unit has a connection to every hidden unit, and every hidden unit has a connection to every output unit. It is called a feedforward network because the direction of connections proceeds from input to output only. There are no connections between units at the same level, and there are no stimulating connections from later units to earlier ones. Each unit is identical, and there is no place for central processing.

与编程的符号系统不同，神经网络能够学习或被训练。虽然技术上存在多种不同的训练/学习机制，但它们基本遵循以下模式：首先，将所有连接的类型（抑制性或兴奋性）和权重（-1 到 0 或 0 到 1）随机化。然后，创建并部署一个已知案例的训练集；对于该网络，它将由已知反射声纳信号的水雷和水下岩石组成。对于我们的探雷网络，输入层代表每个物体的声频谱图样本。

Unlike symbol systems that are programmed, neural nets learn or are trained. Although there are technically several different training/learning regimes, they basically follow this pattern. You begin by simply randomizing all connections with regard to type (inhibitory or excitatory) and weight (-1 to 0 or 0 to 1). Then you create and deploy a training set of known cases; for this network, it will consist of mines and underwater rocks whose reflective sonar signals are known. For our mine detector network, the input layer represents a sample of the sound frequency spectrogram of each object.

由于其随机连接，网络在训练集的几乎所有成员上都会表现不佳。学习通过一种称为反向传播的方法进行，该方法使用输出端的误差来调整网络中单元之间的连接值。因此，你每次遍历网络的一个连接，将每个连接的权重改变一些预先选择的值。例如，在给定的训练运行中，每个权重可能会改变 0.1，从初始随机值 4.3 变为 4.4，或从 -6.6 变为 -6.5。如果给定的更改使网络的选择朝着正确的方向移动（假设声纳剖面是地雷），则保留调整。如果网络朝着错误的方向移动，则撤消调整，连接强度保留其先前的值。在继续进行下一个项目之前，对网络中的每个连接都执行此操作训练集。重复该过程，直到网络学会区分已知矿山和岩石，或者直到它未能满足成功标准。在这种情况下，您需要修改网络，例如添加更多输入单元或隐藏层单元，或者添加额外的隐藏层。

Because of its random connections, the network will do badly on almost all members of the training set. Learning proceeds via a method called backpropagation, by which error at the output end is used to tweak the values of the connections between the units in the network. Thus, you go through the network one connection at a time, altering each connection weight by some pre-chosen value. For example, on a given training run, each weight might be altered by 0.1, going from an initial random 4.3 to 4.4, or from -6.6 to -6.5. If a given change moves the network’s choice in the correct direction (let’s assume the sonar profile is of a mine), the tweak is retained. If the network moves toward an error, the tweak is undone, and the connection strength is left at its prior value. You do this for every connection in the network before proceeding to the next item in the training set. You repeat the process until the network has learned to distinguish known mines from rocks, or until it fails to meet your criteria for success, in which case you need to alter the network perhaps by adding more input units or units in the hidden layer, or perhaps by adding additional hidden layers.

联结主义有时被称为认知科学中的亚符号范式。神经网络不包含诸如Fonebone或barbeque之类的符号，也不包含用于形式计算的抽象符号。然而，个体单元或单元组在与世界互动时，仍有可能自发地获得表征能力。

Connectionism is sometimes referred to as the subsymbolic paradigm in cognitive science. Neural networks don’t contain symbols, such as Fonebone or barbeque, nor abstract tokens of them for the purposes of formal computation. However, there remains the possibility that individual units or groups of units may spontaneously acquire representational capacities as they interact with the world.

网络过程仅发生在大脑和神经系统中，就像符号系统架构一样。认知的符号系统架构提出了这样一种可能性：如果你的个性计算机程序能够被编译，它就可以被下载并在云端实现永生。联结主义者提出了一种云端永生的​​“神经”变体，即构建一个完整的大脑模拟网络并将其上传（Koene & Deca，2013）。

Network processes occur only in the brain and nervous system, as in the symbol-system architecture. The symbol-system architecture of cognition holds out the possibility that if the computer program of your personality could be compiled, it could be downloaded and achieve immortality in the cloud. Connectionists propose a “neural” variation of immortality in the cloud via building a whole brain emulation network and uploading it (Koene & Deca, 2013).

类似地，消除主义者应该相信，将大脑从身体中移除，如果大脑的感觉输入和运动输出被适当地连接起来，模拟你所感知的世界和你在其中采取的行动，你的生命似乎就能继续下去，从而实现永生 (Leahey, 2017)。各种俗气的科幻电影都呈现了这种场景（例如，“三曲腿图的游戏者”，^{2 这}是原版《星际迷航》第二季中的一集，1968 年 1 月 5 日。在那一集中，缸中之脑——一个失落已久的种族的最后幸存者——用一种叫做 Quatloos 的货币，押注人类和类人战士之间角斗士比赛的结果。）最近，《黑客帝国^3》系列电影探索了类似的领域，只不过不仅大脑，而且身体也被储存在缸中，但实际上生活在一个计算机生成的世界中，即同名的矩阵中。

In a similar vein, eliminativists should believe in achieving immortality by your brain being removed from your body and your life seemingly continued if your brain’s sensory inputs and motor outputs were suitably hooked up to simulate the world you sense and the actions you take in it (Leahey, 2017). Various cheesy sci-fi films have presented this scenario (e.g., “The Gamesters of Triskelion,”² an episode from the second season of the original Star Trek, January 5, 1968. In that episode, brains in vats—the last survivors of a long-lost race—placed bets among themselves in a currency called Quatloos on the outcomes of gladiatorial fights among human and humanoid combatants.) More recently, The Matrix³ movies explored similar territory, except that not just brains, but also bodies were stored in vats but lived virtually in a computer-generated world, the eponymous matrix.

预测处理范式

The Predictive Processing Paradigm

古典认识论未能解决的问题是，在追求真理的过程中，如何修正信念。这并非柏拉图等理性主义者所关注的问题。对他们而言，真理与谬误之间的界限十分清晰：一个断言当且仅当其能够被正式证明时，才为真。然而，对于经验主义者而言，真理建立在证据之上，将观察结果与真理的候选者联系起来是一个鲜活而关键的问题，它既包含规范认识论的层面，也包含描述心理学的层面。在认识论中，问题在于我们应该如何根据证据修正信念；在心理学中，问题在于我们如何根据证据修正信念。休谟对这两个问题都进行了探讨，他认为，我们学习真理的方式是逐渐联想观念，而这一过程是动物心智运作的基础。然而，休谟的答案并未令其他思想家满意，他们认为这并不能很好地回答规范性问题。经验和联想观念提供了有用的信念，但他们担心，这些信念并非真正的信念。

Unaddressed by classical epistemologies was how to revise beliefs as we pursue Truth. This was not an issue for rationalists such as Plato. For them, the boundary between truth and error was sharp: An assertion was true if and only if it was formally provable. For empiricists, however, for whom truth rested on evidence, connecting observations to candidates for truth was a live and critical issue, having both normative epistemological and descriptive psychological aspects. In epistemology, the question was how we ought to revise beliefs in light of evidence; in psychology, the question was how we do revise beliefs in light of evidence. Hume addressed both questions, proposing that gradual association of ideas is how we learn truths and that this process was basic to the operation of animal minds. However, Hume’s answer was unsatisfactory to other thinkers, who found it a poor answer to the normative question. Experience and association of ideas provided useful beliefs, but not, they feared, True ones.

与休谟同时代的英国牧师托马斯·贝叶斯（1701-1761）曾探讨过这个规范性问题，但他的解决方案——贝叶斯定理——是在他死后发表的（Bayes，1763），并且被搁置了数十年。贝叶斯定理描述了如何在给定证据的情况下修正对某个信念的信心。A 表示信念本身；“先验”A 表示在收集证据之前对信念 A 的信心；后验 A 表示在给定证据的情况下对信念 A 的修正信心。新证据；B 是证据；p(B) 是证据的概率；p(X|Y) 是条件概率：在 Y 为真的情况下 X 为真的概率^。9

A contemporary of Hume’s, the English minister Thomas Bayes (1701–1761) tackled the normative problem, but his solution, Bayes’ theorem, was published posthumously (Bayes, 1763), and it lay fallow for decades. Bayes’ theorem describes how to revise confidence in a belief given evidence for it. A is the belief; “Prior” A is confidence in belief A before collecting evidence; Posterior A is the revised confidence in belief A given the new evidence; B is the evidence; p(B) is probability of the evidence; p(X|Y) is conditional probability: the probability that X is true given that Y is true.⁹

休谟（1777年，第87页）在论述奇迹的存在时写道：“智者会根据证据调整自己的信念。”贝叶斯定理规定了如何进行调整。休谟和志同道合的怀疑论者认为人类是轻信的，而本书的关键教训是，他们是正确的。我们过度相信自己的感官（第五章），过度相信自己的记忆（第六章），过度相信自己的推理（第七章），因为我们的大脑进化出了神经计算技巧（第八章），用来解决生存问题（第九章），而不是为了探寻真相。

In addressing the existence of miracles, Hume (1777, p. 87) wrote, “A wise man proportions his belief to the evidence.” Bayes’ theorem prescribes how to do the proportioning. Hume and like-minded skeptics think that humans are credulous, and the key takeaway lesson of this book is that they are right. We trust our senses more than we should (Chapter 5), our memory more than we should (Chapter 6), and our reasoning more than we should (Chapter 7) because our brains evolved neural computing tricks (Chapter 8) to solve problems of survival (Chapter 9), not to Know the Truth.

休谟的怀疑论针对的是宗教奇迹，他提出了欺骗的可能性。自然法则被违反的可能性更大，还是人们欺骗的可能性更大？p (B) 的值应该较低。魔术师会欺骗——这就是魔术的定义——并且他们知道这很容易。令人惊讶的是，科学家本应是批判性思考者，但魔术师却发现他们很容易被观众会嗤之以鼻的戏法所愚弄。10^贝叶斯定理告诉你如何利用你先前的信念值和证据的可能性来修正你对奇迹或任何真理主张的信念。

Hume’s skepticism was directed at religious miracles, and he raised the possibility of deception. Is it more probable that the laws of nature be violated or that people might deceive? The value of p(B) should be low. Magicians deceive—it’s the definition of conjuring—and are aware how easy it is. Surprisingly, scientists are supposed to be critical thinkers, but magicians find that they are easily fooled by tricks at which our audiences would hoot.¹⁰ Bayes’ theorem then tells you how to use the values of your prior belief and the likelihood of the evidence to revise your belief in miracles or any claim to truth.

随着贝叶斯定理在 21 世纪变得越来越有影响力（部分原因是作为对复制危机的解决方案），它在神经科学中催生了贝叶斯大脑的概念（Knil & Pouget，2004；Hutchinson & Barrett，2019），并成为预测处理范式 (PPP)。

As Bayes’ theorem became more influential in the 21st century—partly as a solution to the replication crisis—it gave rise in neuroscience to the concept of the Bayesian brain (Knil & Pouget, 2004; Hutchinson & Barrett, 2019), which became the predictive processing paradigm (PPP).

预测处理范式提供了一个神经科学解释，解释了大脑如何以及为何自组织（Friston，2018，2019；Rao & Ballard，1999），并且与联结主义兼容。尽管神经系统不使用反向传播进行学习，但最近的研究表明，神经学习机制在数学上等同于反向传播（Millidge、Tschantz & Buckley，2020；Sacramento、Costa、Bengio & Senn，2018；Whittington & Bogacz，2017，2019）。预测处理（在人工智能中称为预测编码）的总体思路是，我们的大脑构建一个不断更新的世界模型。实时地，并且模型就是意识本身（Clark，2008、2013、2016）。从某种意义上说，PPP 声称意识是一种受控的幻觉。在传统观点中，意识反映现实，要么直接（现实主义），要么通过表征间接（经验主义）。根据 PPP 的说法，意识反映了我们对世界的期望，只有当从感知中接收到错误信息时，这种“幻觉”才会得到纠正。PPP 将某些精神障碍（Adams、Brown 和 Friston，2013）和某些药物的影响（Carhart-Harris 和 Friston，2019）解释为错误纠正出错的场合，从而为它们提供了有趣的启示（图 3.6）。

The predictive processing paradigm provides a neuroscience account of how and why brains self-organize (Friston, 2018, 2019; Rao & Ballard, 1999) that is compatible with connectionism. Although nervous systems don’t use back-propagation to learn, recent work has shown that neural learning mechanisms are mathematically equivalent to back-prop (Millidge, Tschantz, & Buckley, 2020; Sacramento, Costa, Bengio, & Senn, 2018; Whittington & Bogacz, 2017, 2019). The general idea of predictive processing (called predictive coding in AI) is our brain builds a model of the world that is fluidly updated in real time and that the model is consciousness itself (Clark, 2008, 2013, 2016). In a sense, consciousness, PPP claims, is a controlled hallucination. In conventional views, consciousness reflects reality, either directly (realism) or indirectly via representations (empiricism). According to PPP, consciousness reflects our expectations about the world, and this “hallucination” is corrected only when error messages are received from perception. PPP throws interesting light on some mental disorders (Adams, Brown, & Friston, 2013) and the effects of some drugs (Carhart-Harris & Friston, 2019) by explaining them as occasions when error correction goes awry (Figure 3.6).

行为不仅被视为（甚至主要被视为）获取资源的一种手段，也被视为改变世界和/或自身在其中的位置的一种方式，以便提取用于更新模型的信息。强调行动作为信息收集，将 PPP 与我们稍后将讨论的具身认知范式联系起来（Badcock 等人，2019）。

Behavior is seen not only, or even primarily, as a means of obtaining resources, but also as a way of changing the world and/or one’s position in it in order to extract information used to update the model. Emphasis on action as information gathering ties PPP to the embodied cognition paradigm we’ll discuss later (Badcock et al., 2019).

一个思维/大脑中的混合架构

Hybrid Architectures in One Mind/Brain

从20世纪50年代到21世纪初，联结主义被认为是心理学和纯人工智能领域中符号系统认知架构的替代方案。然而，在心理学领域，人类心智的双系统观应运而生，并吸收了这两种架构（Marcus，2019），人工智能领域也出现了类似的融合（Laird、Lebiere & Rosenbloom，2017；Garnelo & Shanahan，2019）。纯人工智能中混合系统的一个实际案例是IBM的“辩论者计划”（Project Debater）（Slonim et al.，2021），它在某种程度上是一个符号系统，其参数由神经网络解析。联结主义解释了系统1中发生的不透明的、哺乳动物的、无意识的信息处理，而符号系统架构则解释了系统2中缓慢的串行处理和使用表征的操作。

From the 1950s into the 2000s, connectionism was presented as an alternative to the symbol-system architecture of cognition in both psychology and pure AI. However, in psychology, the two-system view of the human mind emerged and absorbed both architectures (Marcus, 2019), and a similar rapprochement took place in AI (Laird, Lebiere, & Rosenbloom, 2017; Garnelo & Shanahan, 2019). A practical example of a hybrid system in pure AI is IBM’s Project Debater (Slonim et al., 2021), which is, in part, a symbol system whose arguments are parsed by neural nets. Connectionism provides an account of the opaque, mammalian, unconscious information processing taking place in System 1, and the symbol-system architecture explains the operations of the slow serial processing, representation-using System 2.

建筑 III：激进行为主义

Architecture III: Radical Behaviorism

激进行为主义的创始人是伯勒斯·弗雷德里克·斯金纳（1904-1990）。他的心理学体系包含三个部分：首先，一种哲学，即激进行为主义，尽管其形成早于激进行为主义，但它是具身认知的一种认知架构（Barrett，2016；Morgan，2018）；其次，一套关于行为的理论和研究体系，主要以动物实验为主，即对行为的实验分析；第三，一个在实验室之外审视人类行为的视角，即对人类行为的解读。

The creator of radical behaviorism was Burrhus Frederic Skinner (1904–1990). His scheme of psychology comprised three parts: First, a philosophy, radical behaviorism, which although formulated before it, is a version of embodied cognition’s architecture of cognition (Barrett, 2016; Morgan, 2018); second, a body of theory and research on behavior, primarily with animals, the experimental analysis of behavior, and third, a lens through which to examine human behavior outside the laboratory, the interpretation of human behavior.

^{尽管斯金纳对其思想发展受到的外界}影响保持沉默，但他受到了哲学现实主义复兴 (Smith, 1986) 的影响，并得出了一个真正激进的结论：根本就不存在精神生活。我们可以用斯金纳对精神分析的批判 (Skinner, 1954) 来说明这一观点。弗洛伊德的吸引力很大程度上在于他对人类心灵的华丽描绘，以及其深邃的无意识过程 (Crews, 2017)。斯金纳认为，弗洛伊德做出了一个伟大的发现，但随后犯了一个大错误。这个伟大的发现是，意识不知道行为的原因，因此也不是行为的原因之一；这个大错误是用无意识的（不是错误）、不必要的（才是错误）心理机制填补了原因和结果之间的时间空白。

Although he was silent about outside influences¹¹ on the development of his thought, Skinner was affected by a revival of philosophical realism (Smith, 1986), taking it to a genuinely radical conclusion: There is no such thing as mental life. We can illustrate this idea with Skinner’s critique of psychoanalysis (Skinner, 1954). Much of Freud’s appeal was his lush picture of the human mind, with its vast depths of unconscious processes (Crews, 2017). Skinner argued that Freud made one great discovery but then committed one great blunder. The great discovery was that consciousness was ignorant of, and therefore was not one of, the causes of behavior; the great blunder was filling in the temporal gaps between causes and effects with unconscious (not the blunder), unnecessary (that’s the blunder) mental mechanisms.

想象一下，你因参加学业考试而感到极度焦虑，这导致你的成绩下降。精神分析师可能会对你的问题做出如图3.7所示的解释。

Imagine that you are suffering from acute anxiety about taking academic tests that is causing your grades to suffer. A psychoanalyst might give an account of your problem as visualised in Figure 3.7.

随着治疗的进行，我们会发现，你考试焦虑的根源在于童年早期，可能与如厕训练有关。如厕训练至关重要，因为它是孩子行为满足道德权威要求的时刻。具体来说，孩子必须学会按照父母设定的时间表，而不是孩子自己设定的时间表，从自身释放一些东西。考试就是这样！当你不知道答案时，你会怎么做？你个混蛋！你没有意识到这种联系，因为你压抑了如厕训练的经历。然而，这些记忆仍然存在于你的超我中，即人类人格的道德部分。它吸收了你父母过高的标准，并将他们童年时期强加于你的羞耻感和内疚感强加于你的自我，但由于压抑，你体验到的是对考试的焦虑。一旦我们揭示了你的童年与现在之间的联系，你的自我就会变得更加强大，应该能够应对超我的要求。

We will find as we work through therapy that the cause of your test anxiety is in early childhood, probably having to do with toilet training, an important event because it’s when a child’s behavior meets the demands of moral authorities. Specifically, the child must learn to yield up something from within itself on a schedule set by parents, not the child. And that is what a test is! What do you do when you don’t know the answer: You bullshit! You are not aware of this connection because you have repressed the experience of toilet training. Nevertheless, the memories still live in your superego, the moral part of human personality. It has assimilated your parents’ excessively high standards, and it inflicts on your ego the feelings of shame and guilt they inflicted on you as a child, but because of the repression, you experience them as anxiety about tests. Once we reveal the connections between your childhood and today, your ego will become stronger and should be able to cope with the demands of the superego.

行为分析治疗师的故事更简单（参见图 3.8）：

A behavior analytic therapist has a simpler story (see also Figure 3.8):

你现在的行为，即考试焦虑，是童年时期严格的如厕训练造成的。句号。引入“压抑”（没人能看到）、“超我”（没人能看到）或“自我”（没人能看到）都无法解释或纠正你的考试焦虑。你的父母可能在如厕训练期间错误地使用了负强化和惩罚——疼痛控制方法。疼痛控制会导致恐惧和攻击等不良情绪副产品，而这些副产品现在会在类似情况下再次出现，我们称之为刺激泛化。我们将使用行为方法逐渐消除不愉快的情绪，并强化备考和成功通过考试的替代策略。

Your current behavior, test anxiety, was caused by severe toilet training in childhood. Full Stop. Throwing in “repression” (which no one can see), the “superego” (which no one can see), or the “ego” (which no one can see) adds nothing to an explanation of, or a method of remediation for, your test anxiety. Your parents probably wrongly used negative reinforcement and punishment—pain control methods—during toilet training. Pain control causes unfortunate emotional by-products such as fear and aggression, and these now recur in similar situations, by what we call stimulus generalization. We will use behavioral methods to gradually extinguish unpleasant emotions and reinforce alternative strategies for preparing for and taking tests successfully.

斯金纳的批判不仅适用于精神分析，也适用于那些带有操作性术语的理论，例如赫尔的行为主义（Skinner，1950，1987）。赫尔的核心行为假设（参见第四章）是：有效反应潜能=习惯强度x驱动力，并且每个术语都是操作性定义的。然而，如果我们用操作性定义代替概念，我们就可以消除对理论术语的引用：潜伏期反应是强化反应次数和食物匮乏小时数的函数（l = nxh）。句号。同样，符号系统认知科学认为，生物体计算规则是因为它们的行为与规则一致。然而，物理学家不会说空间中运动的物体计算牛顿运动定律，因为它们的行为与这些定律一致。物体的运动受牛顿定律支配，但它们并不遵循（计算）牛顿定律。同样，生物体受各种行为定律支配，例如泛化，但它们并不遵循，即计算这些定律（Skinner，1966 年、1985 年；另见 Smolensky，1988 年）。

Skinner’s critique applies not only to psychoanalysis, but also to theories with operationally defined terms, such as Hull’s behaviorism (Skinner, 1950, 1987). Hull’s central behavior postulate (see Chapter 4) was: Effective reaction potential = Habit strength x Drive, and each term was operationally defined. However, if we substitute the operational definitions for the constructs, we can eliminate references to theoretical terms: latency of response is a function of number of reinforced responses and number of hours of food privation (l = n x h). Full Stop. Similarly, symbol-system cognitive science says that organisms compute rules because their behavior is consistent with them. However, physicists do not say that objects moving in space compute Newton’s laws of motion because their behavior is consistent with those laws. Objects’ movements are governed by Newton’s laws, but they do not follow (compute) Newton’s laws. Similarly, organisms are governed by various laws of behavior, such as generalization, but they do not follow, i.e., compute them (Skinner, 1966, 1985; see also Smolensky, 1988).

激进行为主义的一个不寻常之处是它对控制环境的定义。直觉上，“我”和“环境”之间的界限是皮肤，心理学家们也默认了同样的观点。但激进行为主义者认为，皮肤并不是一个特殊的屏障：皮肤里面的东西是导致你行为的环境的一部分。例如，牙痛和噪音一样，都是物理刺激；它源于身体，这使得它成为控制环境的一部分。对激进行为主义者来说，“我”就是我所做的，而不是位于身体或大脑中导致行为的内在事物。你的身份就是你的行为，句号。

An unusual aspect of radical behaviorism is its definition of the controlling environment. Intuitively the boundary between “me” and “the environment” is the skin, and psychologists have tacitly adopted the same view. But radical behaviorists say that the skin is not a privileged barrier: what’s inside it is part of the environment that causes your behavior. A toothache, for example, is just as much a physical stimulus as is a noise; its origin in the body makes it part of the controlling environment. For radical behaviorists, “me” is what I do, not an inner thing located in the body or brain that causes behavior. Your identity is your behavior, Full Stop.

由于具身认知和激进行为主义都植根于感知实在论，斯金纳堪称具身认知运动的先驱。“人们被自身经历所改变；他们不会将经验的副本储存为表征或规则。”（斯金纳，1985，第300页）

Because both embodied cognition and radical behaviorism are rooted in perceptual realism, Skinner was a forerunner of the embodied cognition movement. “People are changed by their experiences; they do not store copies of them as representations or rules” (Skinner, 1985, p. 300).

蓄电池比信息处理更能体现行为生物体。我们把电放进电池里，需要的时候再取出来，但电池里并没有电。当我们“放电”的时候，其实是在更换电池，而当它被打开时，“放出电”的其实是一块更换过的电池。

（Skinner，1985，第 294-295 页）

A storage battery would be a better model of a behaving organism [than information processing]. We put electricity into a battery and take it out when needed, but there is no electricity in the battery. When we ‘put electricity in’ we change the battery, and it is a changed battery that ‘puts out electricity’ when tapped.

(Skinner, 1985, pp. 294–295)

斯金纳（1977，第6页）在他最后一篇论文《为什么我不是认知心理学家》中写道：“身体在接触点对世界做出反应；复制是浪费时间。” 他拒绝接受生理学解释，而强调身体活动。“心智是身体所做的事情。它是人所做的事情。……专注于[大脑]就像重拾荷马史诗中的希腊人”（斯金纳，1987，第784页）。“观察一个物体……不仅仅是被动感知；它是一种行为”（斯金纳，1954/1972，第190页）。

Skinner (1977, p. 6) wrote in his last paper, “Why I Am Not a Cognitive Psychologist,” “The body responds to the world, at the point of contact; making copies would be a waste of time.” He rejected physiological explanation for emphasis on bodily activity. “The mind is what the body does. It is what the person does. … To focus on [the brain] is to rejoin the Homeric Greeks” (Skinner, 1987, p. 784). “To see an object … is not mere passive sensing; it is an act” (Skinner, 1954/1972, p. 190).

建筑学 IV：具身认知

Architecture IV: Embodied Cognition

从 19 世纪末认知科学和现象学的早期（Dreyfus & Dreyfus, 1990; Ratcliffe, 2007）到当代机器人技术（Brooks, 1991），这是一种另类架构，直到最近才被认定为符号系统或联结主义架构的替代方案。并非所有拥护者都认同其最常用的名称——具身认知，或坚持其所有论点——其中既有激进派（Chemero，2009），也有温和派（Clark，2008）——但他们提出了一些值得思考的重要观点（Roth & Jornet，2013）。

From the early days of cognitive science and phenomenology in the late 19th century (Dreyfus & Dreyfus, 1990; Ratcliffe, 2007) to contemporary robotics (Brooks, 1991), there was an alternative architecture that, until recently, never jelled as an alternative to the symbol-system or connectionist architectures. Not all its adherents sign on to its most-used name, embodied cognition, or adhere to all its theses—there are radicals (Chemero, 2009) and moderates (Clark, 2008)—but they make important points worth considering (Roth & Jornet, 2013).

想象一下，你被困在一家着火的酒店房间里。你无法从门口离开，因为火势已经蔓延到大厅。这就带来了一个实时问题——如何在火势吞噬你的房间之前逃生——以及框架问题——你和你所处环境的哪些特征与你的逃生相关？

Imagine that you are trapped in a room of a hotel on fire. You can’t leave by the door because the fire has reached the hall. This poses a real-time problem—escaping before the fire consumes your room—and frame problem—what features of you and your environment are relevant to effecting your escape?

你注意到你的房间有一扇窗户（虽然打不开），并想既然你在三楼，或许可以打破窗户，跳到地面后还能活下来。怎么打破窗户呢？需要注意的是，你并没有被编程（符号系统或许会被这样编程）或训练（联结网络或许会被这样训练）来完成打破一大块牢牢固定住的玻璃的任务。

You observe that your room has a window (though it can’t be opened), and figure that since you are on the third floor, you might be able to break the window and survive a jump to the ground. How to break the window? It’s important to note that you have not been programmed (as a symbol system might have been) or trained (as a connectionist network might have been) for the task of breaking a large rigidly held pane of glass.

你环顾四周，想找点东西砸碎窗户。像免费洗漱用品这样的东西显然是错的，所以你从来没想过。这个“结论”看起来很傻，但从具身视角来看，它很有启发性，因为它根本不是一个“结论”。你不会考虑每件洗漱用品的重量，计算 F = M × A，就得出结论说它打不破窗户。同样，其他物品由于不同的原因也没有浮现在脑海中。梳妆台抽屉太重了；钢制床架也许可以当撞锤，但是没有时间拆开床，折叠床架，然后冲向窗户。排除了这么多选择后，一些选择进入了你的脑海，比如椅子或木脚凳：不要太重，不要太轻，而且可以扔。你的身体状况也可能会影响你的选择。 30岁的人可能会选择结实的办公椅，而像我这样75岁的人，可能会选择脚凳——我可以抬起脚凳，但抬不起椅子。假设你我都能做出明智的选择，并且安然度过。

You look around for something to smash the window. Things like the complimentary toiletries are so obviously wrong they never enter your mind. This “conclusion” looks silly, but from the embodied perspective it’s instructive because it is not a “conclusion” at all. You don’t think about each toiletry’s weight, compute F = M × A, and conclude that it can’t break the window. Similarly, other items don’t rise to mind for different reasons. The dresser drawer set is too heavy; the steel bedframe might make a viable ram, but there’s no time to disassemble the bed, fold the frame up, and make a run at the window. With so many choices ruled out, a few enter thought, such as a chair or a wooden footstool: not too heavy, not too light, and throwable. The condition of your body may sway your choice, too. A 30-year-old might go for the solid desk chair, while the same person at age 75, like me, might settle for the footstool—I can lift it, but not the chair. Let’s assume you and I choose wisely and survive.

具身认知运动的目标是笛卡尔式的“侏儒”（homunculus），该理论认为，每个行为生物体内都存在一个智能设备或场所，负责执行智能行为所需的计算。然而，该运动的成员对此观点可能存在分歧，他们认为智能是生物体与其环境之间的关系。智能并非存在于身体中（例如侏儒），而是存在于身体在其所处情境中的行为中。具身认知的另一个名称是“情境认知”。

The target of the embodied cognition movement is the Cartesian homunculus, the thesis that somewhere within every behaving organism there’s an intelligent device or place that performs computations responsible for intelligent behavior. Instead, a point on which members of the movement may differ, intelligence is a relationship between an organism and its environment. Intelligence is not in the body—a homunculus—it is of the body acting in the situation it’s in. One alternative name for embodied cognition is situated cognition.

具身认知的种类

Varieties of Embodied Cognition

情境认知：罗德尼·布鲁克斯的机器人技术

布鲁克斯制造的机器人不使用符号系统或联结主义方法，因为他拒绝表征，而倾向于感知现实主义。他从模仿昆虫的简单机器人——动物机器人开始。昆虫是进化过程中的持久幸存者，因此它们堪称自然界的杰作。它们没有中央计算大脑，只有遍布全身的简单神经系统。同样，布鲁克斯的动物也没有中央处理器，只有简单的反射执行器，独立控制着它们的肢体。有序的运动源于每个肢体与环境的独立接触。动物直接适应自身处境，而非适应环境的表象^。14

Brooks builds working robots that don’t use symbol-system or connectionist methods because he rejects representations in favor of perceptual realism. He began with animats, simple robots that imitate insects. Insects are durable survivors of evolution, so they count among nature’s successes. They have no central computing brain, but rather a simple nervous system spread out through their bodies. Likewise, Brooks’ animats don’t have CPUs, but simple reflexive actuators independently controlling their limbs. Organized locomotion emerges out of each limb’s separate contacts with the environment. Animats adapt directly to their situation, not to a representation of the situation.¹⁴

生态心理学：JJ吉布森的生态现实主义

具身认知在心理学中根深蒂固，源于詹姆斯·J·吉布森（1904-1979）的生态现实主义（Gibson，1950、1972、1977、1979）。感知表征理论的动机之一是，人们显然需要利用二维感官线索构建三维世界的内在精神图景，就像文艺复兴时期艺术家使用暗箱一样。一位名叫阿德尔伯特·埃姆斯的验光师建造了一个扭曲的房间——一种扭曲的暗箱，它能使人产生错觉，让人看到生动逼真的深度感知幻象^。15

Embodied cognition has a strong root in psychology, in the ecological realism of James J. Gibson (1904–1979) (Gibson, 1950, 1972, 1977, 1979). One motive for the representational theory of perception was the apparent need to build an inner mental picture of the 3-D world out of 2-D sensory cues, as did Renaissance artists using a camera obscura. An optician named Adelbert Ames built a distorted room—a sort of distorted camera obscura, which fools people into seeing vivid, living illusions of depth perception.¹⁵

但吉布森的实在论者指出，艾姆斯房间并不能代表现实生活中的感知，因为它人为地禁止了身体运动，迫使我们进入暗箱。图3.9展示了艾姆斯房间的设置；它本质上是一个魔术。

But Gibsonian realists point out that the Ames room is not representative of perception in real life because it artificially forbids bodily movement, forcing us into a camera obscura. Figure 3.9 shows the setup of the Ames room; it is basically a magic trick.

艾姆斯错觉之所以有效，是因为观察者必须透过一个小小的窥视孔才能看到。只要观察者稍稍移动一下，错觉就会消失。由于身体无法活动，我们被迫进入笛卡尔剧场的暗箱，并被迫得出这样的结论：三维意识世界是用户构建的错觉（丹尼特，1991）。拥有实体的生物可以自由探索自己的世界，直接体验深度。

The Ames illusion works only because the viewer must look through a little peephole. If the viewer can move even a little, the illusion vanishes. Denied the mobility of our bodies, we are forced into the camera obscura of the Cartesian Theater and pushed toward the conclusion that the world of 3-D consciousness is a constructed user illusion (Dennett, 1991). Free to explore their world, embodied creatures experience depth directly.

吉布森为具身认知贡献了可供性的概念。传统的感知观点强调我们如何处理和理解物体的感知品质，例如颜色、大小和形状。吉布森拒绝了这种绘画式的、观察性的智力表征，坚持认为感知是为了行动而不是表征。吉布森认为，我们直接感知的不仅仅是物体的感官属性，还有我们可以用物体做什么——它能给我们带来什么动作——并且这些可供性对于不同身体类型的生物来说会有所不同。例如，一块石头可以为老鼠提供藏身或攀爬的地方，为猫提供寻找小猎物藏身的地方，而一个人可以提供一个可以扔出或用作锤子的工具。

Gibson contributed to embodied cognition the concept of affordance. Traditional views of perception emphasize how we process and understand an object’s perceptual qualities, such as color, size, and shape. Gibson rejected this painterly, spectatorial characterization of intelligence, insisting that perception is for action rather than representation. Gibson argued that we directly perceive not just an object’s sensory properties, but also what we can do with an object—what actions it affords us—and that these affordances will be different for creatures with different bodies. For example, a rock will afford a mouse a place to hide or climb on, a cat a place to find small prey hiding, and a person a tool to throw as a projectile or use as a hammer.

动力系统：瓦特蒸汽机调速器案例研究

范盖尔德 (Van Gelder, 1995, p.345) 问道：“如果不是计算，认知会是什么？”通过对比詹姆斯·瓦特 (James Watt, 1736–1819) 为控制他的新蒸汽机而设计的调速器的计算性和非计算性描述，他给出了答案。

Van Gelder (1995, p. 345) asks, “What might cognition be, if not computation?” He provides an answer by contrasting computational and noncomputational accounts of the governor that James Watt (1736–1819) designed to control his new steam engine.

如果说科学革命建立在世界格局的机械化之上，那么工业革命则建立在劳动的机械化之上，即用机器代替动物或人类从事纯粹的体力劳动。机器工作比动物或人类劳动更高效、更廉价，让人们能够用脑力从事体力要求更低但效率更高的工作。然而，让机器产生能量面临着控制方面的挑战。蒸汽机可以产生巨大的能量，但有效地控制其输出对于平稳运行至关重要。1788年，企业家兼工程师詹姆斯·瓦特解决了这个问题，为蓬勃发展的工业革命做出了重要贡献。他发明了离心式调速器，即瓦特调速器。

If the Scientific Revolution rested upon the mechanization of the world picture, the Industrial Revolution rested upon the mechanization of work, using machines in place of animals or people to do sheer physical labor. Machine work can be done more efficiently and cheaply than animal or human labor, freeing people to do less physically demanding but more productive work with their minds. However, getting machines to produce energy poses challenges of control. A steam engine can make prodigious amounts of energy, but effectively controlling its output is necessary to smooth working. In 1788 entrepreneur and engineer James Watt solved the problem, making an essential contribution to the burgeoning Industrial Revolution. He invented the centrifugal, or Watt, governor.

一根轴连接着两个沉重的球。锅炉内的蒸汽推动轴旋转，蒸汽压力越高，轴旋转得越快，球也旋转得越快。球安装在枢轴上，因此当它们旋转时，离心力会将它们向上推离轴。旋转球的位置控制着进入锅炉的燃料量。快速旋转的球会降低锅炉的温度。随着锅炉温度下降，轴旋转得更慢，球旋转得更少，从而增加了进入锅炉的燃料量，因此温度上升，蒸汽压力上升，轴的转速也随之增加，最终温度降低，如此循环往复。循环，将锅炉的输出维持在稳定的可用水平（van Gelder，1995）（图 3.10）。

There is a shaft attached to two heavy balls. The shaft is spun by the steam in the boiler, and the higher the steam pressure, the faster the shaft rotates, and the faster the balls spin. They are mounted on a pivot, so as they spin centrifugal force pushes them out and up away from the shaft. The position of the spinning balls controls the amount of fuel allowed into the fire heating the boiler. Fast-spinning balls reduce the temperature of the boiler. As the boiler temperature goes down, the shaft rotates more slowly, the balls spin less, increasing the input of fuel into the boiler, so the temperature rises, raising the steam pressure and increasing the speed of the shaft, which lowers the temperature, in a continuous dynamic cycle that maintains the output of the boiler to a steady usable level (van Gelder, 1995) (Figure 3.10).

这件工程学作品与心理学有什么关系？关系很大。需要注意的是，瓦特的新型蒸汽机是一种有目的、目标导向的机器，类似于恒温器。调速器的目标是维持蒸汽机锅炉内安全稳定的压力，它以灵活的方式实现这一目标，根据锅炉温度的变化调节燃料流量。瓦特早在控制论出现一个多世纪前就提出了反馈的概念。

What does this piece of engineering have to do with psychology? Quite a lot. Note that Watt’s new steam engine was a purposive, goal-directed machine, akin to a thermostat. The goal of the governor was to maintain safe and steady pressure in the engine’s boiler, and it pursued this goal in a flexible fashion, adjusting fuel flow as the temperature in the boiler changed. Watt hit on the notion of feedback over a century before cybernetics.

我们可以从人工智能逆向工程的角度来解答瓦特的问题，以此来阐明笛卡尔认知科学与具身认知科学之间的基本区别。让我们运用马尔提出的认知分析层次。

We can illustrate the basic differences between Cartesian cognitive science and embodied cognitive science by addressing Watt’s problem from the reverse engineering point of view of AI. Let’s apply Marr’s levels of cognitive analysis.

• 认知层面：我们如何才能保持发动机稳定、可控的输出？
• The Cognitive Level: How can we maintain steady, controlled output from the engine?
• 算法级别：可以使用哪些规则来管理维持稳定、受控输出的系统？
• The Algorithm Level: What rules might be used to govern a system that maintains steady, controlled output?
• 我们可以构建或使用什么设备来执行这些规则？
• What device can we build or employ to carry out these rules?

接下来，我们进行如下操作。在锅炉上安装一个传感器，每十秒采样一次锅炉内部压力。然后，该传感器将输入提供给以下计算程序（BASIC 语言）：

We then proceed as follows. Put a sensor on the boiler that samples the boiler’s internal pressure every ten seconds. This sensor then provides input to the following computational program (in BASIC):

• 10.输入（压力）
• 10. Input (Pressure)
• 20. 如果P > 500 psi，则转至 40
• 20. If P > 500 psi then goto 40
• 30. 如果P < 450 psi，则转至 50
• 30. If P < 450 psi then goto 50
• 40.减少燃料投入
• 40. Decrease fuel input
• 45. 转到 10
• 45. Goto 10
• 50.增加燃料投入
• 50. Increase fuel input
• 60. 转到 10。
• 60. Goto 10.

然后我们将该程序输入计算机，将计算机连接到锅炉的传感器和燃料控制阀，计算机将执行瓦特调节器的工作，通过调节火的燃料输入来升高或降低温度，从而响应锅炉内的压力。

We then enter this program into a computer, hook the computer up to the sensor and the fuel-control valve for the boiler, and the computer would do the work of the Watt governor, responding to the pressure inside the boiler by regulating the fuel input to the fire to raise or lower the temperature.

在这种情况下，一个通用的信息处理设备将承担瓦特那台不利用信息的专用设备的工作。传统人工智能的假设是，人类是通用的信息处理器，通过学习适当的信息处理步骤来执行任务。

In this scenario, a general-purpose information-processing device would be doing the work of Watt’s non-information-using dedicated device. The assumption of traditional AI is that human beings are general-purpose information processors who carry out tasks by learning the appropriate information-processing steps.

具身认知的倡导者认为，人类更像瓦特的调速器，其智力根植于我们与世界的身体互动。单凭大脑过程——无论是符号系统还是神经网络——不足以解释动物的智力。动物需要一个与世界互动的身体。马尔型系统完全由思维构成，除了提供输入的传感器和服从程序指令的燃油控制阀外，没有其他相关的身体。瓦特的调速器完全由身体构成，没有思维；其控制发动机的行为可以用耦合动态系统的物理定律来解释（van Gelder, 1995）。它不需要表征、规则和计算。生物体可以被视为在动态宇宙中相互作用的连续动力系统。

Advocates of embodied cognition believe that human beings are more like the Watt governor, intelligence being rooted in our bodily interactions with the world. Brain processes alone—whether conceived as symbol systems or as neural nets—do not alone explain animal intelligence. A body that interacts with the world is required. A Marr-type system is all thought with no relevant body except the sensor that provides input and the fuel-control valve that obeys the program’s commands. Watt’s governor is all body and no mind; its behavior in controlling the engine is explained by the physical laws of coupled dynamic systems (van Gelder, 1995). There need be no representations, no rules, and no computations. Organisms may be regarded as continuous dynamical systems interacting with each other in a dynamical universe.

延伸思维：你的思维在你的手机里吗？

具身认知的另一种形式是延展心智论。它认同符号系统假说，认为思维是符号操控，但认为人类的认知系统延伸到大脑之外，进入更广阔的环境中（Clark，2008）。例如，记忆的认知过程通常位于大脑中，但在现代社会，我们的大部分记忆已被转移到智能手机和谷歌等互联网服务中。我们不必记住挚友的电话号码，因为它就在我们手机的数据库中；我们不需要知道《大宪章》是什么，我们可以用谷歌搜索。此外，这种认知过程的转移自人类进化以来就一直在进行：在骨刻、泥板、笔记、操作手册上记录，以及依靠同事了解……关键信息（Hutchins, 1995）。这种记忆卸载解释了为什么现代人类的大脑比第一批智人更小，却仍然能够繁荣发展（Bednarik, 2014）。

Another flavor of embodied cognition is the thesis of the extended mind. It agrees with the symbol-system hypothesis that thinking is symbol manipulation, but it holds that the human cognitive systems extend outside the brain into the larger environment (Clark, 2008). For example, the cognitive process of memory is usually located in the head, but, in the modern world, much of our memory has been offloaded into smartphones and Internet services such as Google. We don’t have to remember our best friend’s phone number because it’s in our phone’s database; we don’t need to know what the Magna Carta was, we can Google it. Moreover, such offloading of cognitive processes has been going on since human beings first evolved: making records on incised bones, clay tablets, notes to oneself, procedural manuals, and relying on coworkers to know critical information (Hutchins, 1995). Such memory offloading explains why modern humans have managed to thrive despite having smaller brains than the first Homo sapiens (Bednarik, 2014).

建筑 V：理性选择理论

Architecture V: Rational Choice Theory

心理学通常被称为社会科学，但由于它关注的是个体心智，因此经常被排除在社会科学书籍之外（例如，Manicas，2006；Leahey，2007）。然而，我们必须纳入主流的社会科学理论——理性选择理论（RCT），因为它在经济决策和行为生态学中可以作为心理学解释的替代（第七章和第九章）。

Psychology is often called a social science, but because it focuses on the individual mind, it is often excluded from books on social science (e.g., Manicas, 2006; Leahey, 2007). However, we must include the dominant social science theory of human behavior, rational choice theory (RCT), because it is an alternative to psychological explanation in economic decision making and behavioral ecology (Chapters 7 and 9).

面对各种各样的选择，从早上穿什么鞋到约谁约会，再到投资什么，人们都会考虑一系列方案，最终选择对自己最有利、效用最大的那个 (Friedman, 1953; Levin & Milgrom, 2004; Lovett, 2006; Gintis, 2018)。该理论的优点在于，它能够构建涵盖人类所有行为的精确数学模型。行为生态学将其应用于动物行为，用适应度（后代数量）代替效用。随机控制理论 (RCT) 与心理学解释（尤其是计算解释）的区别在于，它完全不考虑导致行为的内部过程，无论是心理的还是生理的^。16然而，它与激进行为主义非常相似，都强调量化规律 (Herrnstein, 1990)，甚至与具身认知也一样，这三者都回避了内在机制。

Given a choice between options of any type, from what shoes to wear in the morning to whom to ask for a date to what to invest in, people consider a range of alternatives and choose the one that is best—has the most utility—for them (Friedman, 1953; Levin & Milgrom, 2004; Lovett, 2006; Gintis, 2018). A virtue of the theory is that it allows the construction of precise mathematical models of behavior ranging across all human life. Behavioral ecology applies it to animal behavior, substituting fitness (number of offspring produced) for utility. What sets RCT apart from psychological explanation, especially computational ones, is that it is entirely agnostic about internal processes, mental or physiological,¹⁶ that cause behavior. However, it is much like radical behaviorism, with which it shares an emphasis on quantitative laws (Herrnstein, 1990), and even to embodied cognition, in that all three eschew inner mechanisms.

非计算架构让我们想起认知科学中最根本的问题——心智本身的本质。心智是像原子一样的自然类型，等待着科学的发现；还是像燃素一样的虚构，在科学中暂时有用，但并非真实存在，可以被更好的概念取代？具身认知与激进行为主义非常相似。两者都对感知持实在论观点，并都强调生物体与世界之间的物理相互作用作为适应性行为基础的重要性。两者之间的主要区别在于，具身认知的倡导者假设心智是自然类型，真实存在于自然界中。他们认为笛卡尔对心智的描述是错误的，需要被取代，就像早期关于原子是物质最小单位的观念需要被更复杂的亚原子粒子理论取代一样。另一方面，激进行为主义者认为，笛卡尔受到先前存在的宗教思想的误导，发明了一个虚构的概念——心智，与宙斯或燃素并列。 “心智”这个概念不应该被修改，而应该被抛弃。

Noncomputational architectures remind us of the most fundamental issue in cognitive science, the nature of mind itself. Are minds natural kinds like the atom, awaiting discovery by science, or are minds fictions like phlogiston, useful for a time in science but not real, replaceable by better concepts? Embodied cognition looks a lot like radical behaviorism. Both take a realist view of perception and both stress the importance of physical interaction between the organism and the world as the basis of adaptive behavior. The difference between them is mostly that advocates of embodied cognition assume that minds are natural kinds, really existing in nature. They say that Descartes’s description of mind was wrong and needs replacing, the way that early ideas about atoms as the smallest units of matter needed to be replaced by more sophisticated theories about subatomic particles. Radical behaviorists, on the other hand, think that Descartes, misled by preexisting religious thought, invented a fictitious concept, mind, on a par with Zeus or phlogiston. The concept of “mind” should not be revised but dropped.

我们现在的情况

Where We Are Now

我们现在准备探索人类认知，从其行为的基本构成要素到更高级的心理过程。除了人类思维的双系统图谱之外，我们还将在一个我称之为认知科学理论非还原层级的框架内开展工作。我并没有采用那种每个层级都被降低到（或被消除到）较低层级的层级结构，而是结合了马尔的层级结构，并根据相邻层级理论所回答的解释性问题的类型来定义它们之间的关系。每个层级都回答其上一层级提出的“如何”问题，并向其下一层级提出“为什么”问题（图 3.11）。

We are now prepared to explore human cognition from its basic building blocks of behavior to the higher mental processes. In addition to the two-system map of the human mind, we will work within a framework I call the non-reductive hierarchy of theories in cognitive science. Instead of being a hierarchy in which each level is reduced to (or eliminated by) a lower one, I incorporate Marr’s levels and define the relationship between theories at neighboring levels in terms of the kinds of explanatory questions they answer. Each level answers how questions posed by the level above it and poses why questions to the level below (Figure 3.11).

外卖

Takeaway

古希腊人说“认识你自己”，苏格拉底说真正的知识需要理性的论证。笛卡尔审视内心，在私人理性中找到了真理。第七章将探讨理性在社会中扮演的重要角色。然而，一些研究表明，理性可能会被过度推崇，以至于干扰良好的判断，尤其是在涉及情感的情况下（Wilson，2002）。

The classical Greeks said know thyself, and Socrates said real knowledge required rational justification. Descartes looked inward and found Truth in private reason. In Chapter 7 we will look at the important role being rational plays in our society. However, some research suggests that it’s possible to push rationality too far, to the point that it interferes with good judgment, especially when emotions are involved (Wilson, 2002).

例如，Wilson 等人（1993）进行了一项实验，让参与者根据自己对艺术海报的喜爱程度对其进行评分，或者在对照组中，简单地填写一份信息表。评分参与者被要求解释他们为何对海报进行如此评分。这些海报分为两类：在预测试中，印象派绘画获得了较高的评价；以及评分较低的简单波普艺术海报。两组的参与者都被告知，他们参与者可以带任何喜欢的海报回家，作为参与的小奖励。实验组成员倾向于选择波普艺术海报，并选择其中一张；而对照组成员则更喜欢印象派海报，并带走了这些海报。之后，研究人员对参与者进行了跟踪调查，询问他们对奖励海报的满意度以及是否还保留着这些海报。实验组的参与者对奖励海报的满意度较低，也不太可能保留这些海报。这里有一个例子，欧比旺建议卢克关闭他的目标计算机（系统2），并相信他的直觉^（系统1）是正确的。

For example, Wilson et al. (1993) performed an experiment in which participants rated art posters for how much they liked them, or, in the control, simply filled out an information form. The rating participants were asked to give reasons why they rated the posters as they had, which were of two types, highly rated (in pre-testing) Impressionist paintings or less highly rated simple pop-art type posters. Both groups’ participants were told they could take any poster they liked home as a small reward for participating. Experimental group members tended to report preferring the pop-art posters and chose one of them, but control group members preferred and took the Impressionists. Later, participants were followed up and asked how happy they were with their reward posters and if they still had them. Experimental group participants were less happy with and less likely to have kept their choices than controls. Here we have a case where Obi-Wan’s advice to Luke to turn off his targeting computer (System 2) and trust his intuition¹⁷ (System 1) is right.

阅读材料

Readings

符号系统假说最重要的创始人是赫伯特·西蒙，他的自传是《我人生的模型》（1991）。我曾评论过这本书（Leahey，1993）；https://cdm15960.contentdm.oclc.org/digital/collection/p15960coll21/id/72873/rec/2。我还写了一篇关于他获得诺贝尔奖的文章（Leahey，2003），doi: 10.1037/0003-066X.58.9.753。

The most important creator of the symbol-system hypothesis was Herbert Simon, whose autobiography is Models of My Life (1991). I reviewed it (Leahey, 1993); https://cdm15960.contentdm.oclc.org/digital/collection/p15960coll21/id/72873/rec/2. I also wrote an article about his Nobel Prize (Leahey, 2003), doi: 10.1037/0003-066X.58.9.753.

斯金纳写了三卷本自传，分别是《我的生平》（1976年）、《一个行为主义者的塑造》（1979年）和《后果的问题》（1983年），均由克诺夫出版社出版。他本人也曾表示，这三卷书秉承着他的哲学，最终讲述的是一个非人。有一个基金会（https://www.bfskinner.org/）致力于维护他的记忆，并推广激进行为主义，主要作为一种疗法。想了解行为主义，请阅读《关于行为主义》（纽约：克诺夫出版社，1974年）。

Skinner wrote a three-volume autobiography, Particulars of My Life (1976), The Shaping of a Behaviorist (1979), and A Matter of Consequences (1983), all published by Knopf. True to his philosophy, in the end, the three volumes were, he said, about a non-person. There is a foundation, https://www.bfskinner.org/, that tends his memory and advances radical behaviorism, primarily as a therapy. To learn about behaviorism, read About Behaviorism (New York: Knopf, 1974).

上一章的人工智能阅读材料包含神经网络的背景知识。目前还没有关于具身认知的非技术性书籍，最接近的是安迪·克拉克的《冲浪不确定性》（纽约：牛津大学出版社，2016年），该书也讨论了PPP。

The AI readings from the previous chapter include background on neural nets. There is yet no non-technical book on embodied cognition, the closest being Andy Clark’s Surfing Uncertainty (New York: Oxford University Press, 2016), which also discusses PPP.

PPP 的领导者之一是阿尼尔·塞斯 (Anil Seth)，他的著作《做你自己：意识的新科学》 (纽约：企鹅出版社，2021 年) 是对 PPP 的非技术性解读。他曾在 TED 上发表演讲，网址：https: //www.youtube.com/watch?v=lyu7v7nWzfo 。

One of the leaders of PPP is Anil Seth whose Being You: A New Science of Consciousness (New York: Penguin, 2021) is a non-technical treatment of PPP. He gave a TED talk, https://www.youtube.com/watch?v=lyu7v7nWzfo.

关于揭穿通灵术和其他神秘事件，请参阅詹姆斯·兰迪 (James Randi) 的《Flim-Flam》（纽约：普罗米修斯出版社，1980 年）以及佩恩和泰勒的《胡说八道》 (Penn & Teller's Bullshit ) 的各个片段。

On debunking psychics and other mysteries, see James Randi, Flim-Flam (New York: Prometheus, 1980) and various episodes of Penn & Teller’s Bullshit.

电影

Movies

显然，^《18 录像带谋杀案》是一部适合观看以培养具体/扩展认知的电影。

Apparently¹⁸ Videodrome is the movie to see for embodied/extended cognition.

值得一看的两部纪录片分别是：第一部是关于詹姆斯·兰迪的《诚实的骗子》 （https://www.imdb.com/title/tt2246565/）；第二部是关于人工智能的《人工智能的乐趣》 （https://www.imdb.com/title/tt9053472/）。这部纪录片的结尾展示了人工智能物体感知能力的一个严重缺陷。对图片进行细微的改动，即使人眼无法识别，也可能使人工智能无法识别。这种情况也延伸到了停车标志之类的地方，因此恐怖分子或犯罪分子可以对交通标志进行细微改动，从而对自动驾驶汽车造成干扰，包括目前那些只需极少驾驶员干预即可行驶的自动驾驶汽车。

Two relevant documentaries to watch are, first, on James Randi, An Honest Liar https://www.imdb.com/title/tt2246565/ and, second, on AI, The Joy of AI https://www.imdb.com/title/tt9053472/. This one closes with demonstrations of a serious glitch in AI object perception abilities. Tiny changes to a picture that don’t change it to human eyes can make it unrecognizable to AI. This extends to things like stop signs, so a terrorist or crook could make tiny changes to traffic signs and cause chaos for self-driving cars, including in models now available that drive with minimal driver input.

魔法

Magic

探水术是一种看似奇迹的现象，指用所谓的超感官方法找到看不见的东西。最常见的版本是探水术（或“巫术”），但所谓的通灵者也会探水寻找油和丢失的物品，并通过探水寻找线索来“帮助”警方。魔术师利用人们对探水术的信仰，通常使用新时代风格的“水晶摆”。例如，在一个魔术中，五张扑克牌、超感官知觉卡或“新时代”卡片放在桌子上。趁魔术师不注意，观众选择一张牌：在演示¹⁹中，他通过探水找到它。魔术师回来并展示一个摆锤，将其稳稳地放在每张牌上方。尽管如此，摆锤仍在四张牌上方来回摆动，并在所选牌上方上下摆动。也可以让观众使用摆锤来找到牌。20^一个看似奇迹的现象发生了。

Dowsing is an apparent miracle, identifying something unseen by supposed extra-sensory methods. The most common version is water dowsing (or “witching”), but alleged psychics also dowse for oil and lost objects, and “help” police by dowsing for clues. Magicians exploit belief in dowsing, usually using New Agey “crystal pendulums.” For example, in one trick, five playing cards, ESP cards, or “New Age” cards are laid on a table. While the magician is not looking, a spectator chooses one card: in the demo¹⁹ he then locates it by dowsing. The magician returns and shows a pendulum, which he holds steady over each card. Nevertheless, the pendulum swings back and forth over four cards and up and down over the chosen card. It’s also possible to get the spectator to use the pendulum to find the card.²⁰ An apparent miracle has occurred.

钟摆戏法是预测、行动以及基于感知的纠错反馈紧密结合的一个例子。它依赖于一种被James称为“意念运动”的现象（Wiese & Metzinger，2017；Lawson & Crane，2014）。仅仅想到一个动作，就会自动启动神经系统来执行它。用PPP（感知-行动）的术语来说，你的大脑“预测”某个动作会发生——或者更准确地说，你的世界模型会发生某种变化——然后运动系统就会采取行动，使预测成真。当魔术师将钟摆悬在选定的牌上方时，他只需想象垂直运动——预测它会发生——运动系统就会毫不费力地让它发生。观众也可以这样做。即使他们试图保持钟摆静止，以掩盖钟摆悬在他们牌上方的事实，魔术师的暗示——预测——也会促使运动系统推动钟摆移动。许多所谓的催眠控制都是以同样的方式起作用的。

The pendulum trick is an example of the intimate coupling of prediction, action, and error-correcting feedback from perception of the act. It depends on a phenomenon James called ideo-motor action (Wiese & Metzinger, 2017; Lawson & Crane, 2014). Simply thinking of an action automatically sets in motion the nervous processes to carry it out. In PPP terms, your brain “predicts” that an action will happen—or, more precisely, that your world-model will change in a certain way—and then the motor system acts to make the prediction true. As the magician holds the pendulum over the selected card, he merely thinks about vertical motion—predicts that it will happen—and the motor system makes it happen with minimal effort. Audience members can be made to do the same thing. Even though they try to hold the pendulum still to hide that it’s over their card, the suggestion—the prediction—from the magician causes the motor system to make the pendulum move. Much of so-called hypnotic control works the same way.

进化与比较认知：意识的目的是什么？

Evolution and Comparative Cognition: What Is Consciousness For?

进化创造了比较认知领域，旨在研究人类与动物心智之间的相似与差异、连续与不连续性。笛卡尔范式的一个假设是，心智和意识是共存的；不存在无意识的心理过程，因为不属于灵魂的东西就属于机械的躯体。因此，主导比较认知的首要问题是：意识（心智）是一种适应吗？以及什么样的外在行为体现了内在的心智？笛卡尔自然没有回答第一个问题，但他对第二个问题——语言的使用——给出了答案。

Evolution created the field of comparative cognition, investigating similarities and differences, continuities and discontinuities, between the minds of humans and animals. An assumption of the Cartesian paradigm was that mind and consciousness were co-extensive; there were no unconscious mental processes because what did not belong to the soul belonged to the mechanical body. Hence, the first questions to dominate comparative cognition were: Was consciousness (mind) an adaptation? and What outward behavior manifested a mind within? Descartes had naturally not addressed the first question, but he did have an answer to the second question, use of language.

在达尔文之后，笛卡尔的语言标准显得过于绝对，造成了一条巨大的鸿沟，一边是动物，另一边是人，这种鸿沟与达尔文关于进化是渐进过程的观点不一致。如果人类心智以达尔文的方式进化，那么它必然有更简单的东西进化而来。人们可以调和笛卡尔和达尔文的理论，但代价是不能回答关于心智的第一个问题——意识是一种适应吗？——意识可能只是繁忙的神经系统中电化学活动的副产品，就像电源线中的热量是电流从电源传输到工作电机的副产品一样。我们将在第 8 章更详细地讨论这一主张。

After Darwin, Descartes’s criterion of language seemed too absolute, creating a yawning chasm with animals on one side and people on the other, a gap inconsistent with Darwin’s view of evolution as a gradual process. If human mind evolved in a Darwinian way, it had to have something simpler to evolve from. One could reconcile Descartes and Darwin at the cost of answering the first question about mind—is consciousness an adaptation?—in the negative. It might be the case that consciousness is nothing but a by-product of electro-chemical activity in a busy nervous system, the way heat in a power cord is a by-product of moving electricity from its source to a working motor. We’ll take up this claim in more detail in chapter 8.

然而，詹姆斯（1890）提出了强有力的论据，支持意识（心智）是一种适应性而非进化偶然​​的观点。他指出，没有意识的生物不会在意自己的生死。相反，詹姆斯他认为意识是为了生存而积极奋斗的斗士，这增加了其拥有者获得达尔文式成功的机会。詹姆斯的观点赢得了众多追随者，并促使心理学家们开始研究：如果不是语言，意识的行为表达可能是什么。

However, James (1890) marshalled powerful arguments in favor of the view that consciousness (mind) is an adaptation rather than an evolutionary accident. He pointed out that a creature without consciousness would not care if it lived or died. Instead, James argued that consciousness is an active fighter for the goal of survival, increasing its bearer’s chances of Darwinian success. James’s view won adherents, moving psychologists to work on the question of what might be the behavioral expression of consciousness if it wasn’t language.

学习代替意识

Learning Stands In for Consciousness

最终，人们达成了共识。在达尔文进化论中，物种世世代代地适应着不断变化的环境。个体生物通过改变其行为，即在其一生中不断学习，来适应不断变化的环境。意识——心智——是一种适应性行为，体现在生物体的学习能力中。此外，如果学习是私人心智在公众领域所做的事情，那么基于主观内省的研究就可以被基于客观学习观察的研究所取代。认知科学家可以在实验室中研究学习是如何进行的，应用心理学家可以运用关于学习的科学发现来帮助人们适应工业现代主义的新世界（杜威，1900/1978）。

Eventually a consensus was reached. In Darwinian evolution, species adapted to their changing environments over generations. Individual organisms adapted to their changing environments by changing their behavior, that is, by learning in the course of their lifetimes. Consciousness—mind—was an adaption that manifested itself in organisms’ capacities to learn. Furthermore, if learning was what private mind publicly does, then research using subjective introspection could be replaced by research using objective observations of learning. Cognitive scientists could study how learning worked in the laboratory, and applied psychologists could apply scientific findings about learning to help people adapt to the new world of industrial modernism (Dewey, 1900/1978).

消除行为主义

Eliminative Behaviorism

在沃森宣言之后，提出了两种用行为取代心灵的方法。一种是卡尔·拉什利 (1890-1958) 的严格行为主义(Lashley, 1923)。拉什利是沃森的学生。他们一起复制了巴甫洛夫的动物实验并发明了人类巴甫洛夫条件反射的方法 (Watson, 1916)。拉什利受到俄罗斯客观心理学还原学派的影响，并将成为美国生理心理学的先驱 (第 8 章)。拉什利的严格行为主义（或更描述性地称生理行为主义）的目标是通过表明它们实际上是大脑过程来消除对心灵（包括意识）的提及，但这一目标超出了拉什利的方法论理解范围。

In the wake of Watson’s manifesto, two ways of replacing mind with behavior were proposed. One was Karl Lashley’s (1890–1958) strict behaviorism (Lashley, 1923). Lashley was a student of Watson’s. Together they replicated Pavlov’s animal experiments and invented methods for Pavlovian conditioning in humans (Watson, 1916). Lashley became influenced by the reductive Russian school of Objective Psychology and would go on to be the American pioneer of physiological psychology (Chapter 8). The goal of Lashley’s strict, or more descriptively, physiological, behaviorism was to eliminate references to mind, including consciousness, by showing that they were really brain processes, but this ambition was beyond Lashley’s methodological grasp.

作为方法论行为主义的计算架构

Computational Architectures as Methodological Behaviorism

行为主义的另一种表述，方法论行为主义（例如，托尔曼，1935，1936），成为了主流心理学和认知科学。它并不承诺解释意识，因为它用一个简单的论证就将意识排除在科学范围之外。科学要求事实是公开的，每个聚集在实验台周围的人都可以获得。然而，内省的“事实”是私密的，锁在每个人的头脑中。方法论行为主义者并没有断言意识是一种幻觉，甚至没有说它不重要，只是它的私密性使它不适合成为科学研究的对象。意识可以留给诗人、画家、小说家以及一般的艺术。科学心理学只有通过研究客观可定义的条件如何规律地改变客观可定义的行为才能作为一门科学发展。方法论行为主义者愿意将未观察到的理论术语纳入他们的理论，只要这些术语能够通过与客观程序或测量联系起来进行精确定义。

The other formulation of behaviorism, methodological behaviorism (e.g., Tolman, 1935, 1936), became mainstream psychology and cognitive science. It did not promise to explain consciousness because it ruled consciousness out of scientific bounds with a simple argument. Science demands facts that are public, available to everyone gathered around a lab bench. The “facts” of introspection, however, were private, locked up in each person’s head. Methodological behaviorists did not assert that consciousness is an illusion, or even that it was unimportant, only that its privacy made it an unsuitable object of scientific study. Consciousness could be left to poets, painters, novelists, and the arts in general. Scientific psychology could only advance as a science by investigating how objectively definable conditions lawfully change objectively definable behaviors. Methodological behaviorists were willing to admit unobserved theoretical terms into their theories provided they could be precisely defined by connecting them to objective procedures or measurements.

将心理学定义为行为科学，可以回避关于意识存在及其本质的棘手争论。计算心理学也是如此。图灵机是通用计算机，神经网络可以进行高性能数学运算，未来机器人或许可以驾驶汽车，但我们不必将意识归因于它们的运作。主流认知科学是方法论行为主义的一种版本（Dennett，1991；Leahey，1994）。

Difficult arguments about the existence and nature of consciousness could be evaded by defining psychology as the science of behavior. So does computational psychology. Turing machines are universal computers, neural nets do high-powered math, and one day robots may drive cars, but we don’t have to attribute consciousness to their operations. Mainstream cognitive science is a version of methodological behaviorism (Dennett, 1991; Leahey, 1994).

发现学习

Discovering Learning

行为主义对认知科学的贡献在于，它对学习过程的细致研究主要基于动物模型，而这受到学习过程相关理论的驱动。当然，使用动物模型是为了方便研究，但受联想主义的影响，行为主义者认为学习规律，即联想规律，至少在所有哺乳动物身上都是相同的。

Behaviorism contributed to cognitive science the careful study of learning, primarily in animal models, driven by rival theories of the learning process. Animal models were used, of course, for convenience, but, influenced by associationism, behaviorists believed that the laws of learning, i.e., of association, would be the same across at least all mammals.

人们提出了各种各样的学习类别。其中一种区别是联想学习和非联想学习。在非联想学习中，反复接触刺激会改变生物体对刺激的反应。在兴奋状态下，反复接触会增强反应；在抑制状态下，反复接触会减弱反应。在联想学习中，我们会在不同条件下向生物体呈现多种刺激，并观察由此导致的行为变化。

Various categories of learning have been proposed. One distinction is between associative and nonassociative learning. In nonassociative learning, repeated exposure to a stimulus changes an organism’s response to the stimulus. In excitation, repeated exposure strengthens the response; in inhibition, repeated exposure weakens the response. In associative learning, multiple stimuli are presented to an organism under various conditions, and we look for changes in behavior that result.

更本质的区别通常体现在试错法（其原名）、工具性学习或操作性学习（斯金纳的术语）与经典学习（其原名）、巴甫洛夫学习（当今的主流术语）或反应性学习（斯金纳的术语）之间。后者是指对某一刺激的特定先天反应（例如恐惧）受到另一个通常不会引发该刺激的刺激的控制。反应（例如，铃声）。前者发生在自由情境中的行为（例如，按下杠杆）以及行为后的后果（例如，允许动物获得食物）会改变该行为的频率。

A more substantial distinction is typically made between trial-and-error (its original name), instrumental, or operant learning (Skinner’s term), on the one hand, and classical (its original name), Pavlovian (the mainstream term today), or respondent (Skinner) learning, on the other. The latter occurs when a given innate response (e.g., fear) to a stimulus is brought under control of another stimulus that does not normally evoke the response (e.g., the sound of a bell). The former occurs when a behavior (e.g., pressing a lever) occurs in a free situation and the consequences (such as allowing the animal access to food), following the behavior change the frequency of that behavior.

试错学习

Trial and Error Learning

美国心理学家爱德华·李·桑代克（Edward Lee Thorndike，1874-1949）发现了试错学习。桑代克致力于研究儿童学习——他的职业生涯都致力于教育心理学——但哈佛大学不允许他这样做，于是桑代克转而研究动物。

American psychologist Edward Lee Thorndike (1874–1949) discovered trial-and-error learning. Thorndike wanted to study learning in children—he spent his career in educational psychology—but Harvard would not permit it, and Thorndike turned to animals.

当时的比较心理学家希望确定不同动物物种的相对智力，始于乔治·罗马尼斯 (1848-1894) 和《动物智力》。罗马尼斯采用了后来被称为轶事法的方法。他收集关于动物行为的故事（轶事），并筛选这些故事，试图重建所涉及动物的心理过程并评估每个物种的智力。在这样做的过程中，罗马尼斯和其他人面临着严重的笛卡尔绊脚石。尽管有局限性，但人们可以报告他们的有意识的想法，而动物不能。轶事法被认为是比较认知中内省法的替代品，但桑代克 (1911) 严厉批评了它。

Comparative psychologists of the time wanted to establish the relative intelligence of different animal species, beginning with George Romanes (1848–1894) and Animal Intelligence. Romanes employed what came to be called the anecdotal method. He collected stories (anecdotes) about the behavior of animals and sifted through the stories attempting to reconstruct the mental processes of the animals involved and assess the intelligence of each species. In doing so, Romanes and others faced a serious Cartesian stumbling block. People can report their conscious thoughts, albeit with limitations, while animals can’t. The anecdotal method was supposed to be the replacement for introspection in comparative cognition, but Thorndike (1911) harshly criticized it.

首先，它高估了动物的心智能力。人们倾向于认为动物很聪明，而不是平凡或愚蠢。1更糟糕^的是，轶事研究法及其对动物思维过程的重构很容易被拟人化。看着动物解决问题，我们很容易设身处地为之，把自己在同样情况下的想法投射到动物身上。最后，因为每个轶事都与一个独特且不受控制的情况有关，所以很难弄清楚哪些环境特征决定了特定动物的行为。桑代克认为实验应该取代轶事。在他的研究方法中，迄今为止最重要的是他所谓的“谜题盒”，因为它们的形式决定了他的学习理论，并将教会我们很多关于方法和理论之间关系的知识。

First, it overestimated animals’ mental powers. People tended to report animals being clever, not being ordinary or stupid.¹ Worse, the anecdotal method, with its attendant reconstruction of the animal’s thought processes, was easily given to anthropomorphism. Watching an animal solve a problem, we easily put ourselves in its place and project onto the animal the thoughts we would have in the same circumstances. Finally, because each anecdote is about a unique and uncontrolled situation, it’s hard to figure out which features of the environment determined a given animal’s behavior. Thorndike argued that experiments should replace anecdotes. Of his methods, by far the most important were his so-called puzzle-boxes, because their form dictated his theory of learning, and will teach us much about the relationship between method and theory.

桑代克制作了几个谜盒（如图 4.1所示），并将一位参与者（通常是一只小猫）放入其中。谜盒是一种笼子，猫可以通过学习操作操纵器从里面打开门。之后，猫逃走了，吃了一些鱼，然后被放回盒子里^。2

Thorndike constructed several puzzle boxes (as shown in Figure 4.1) in which he placed one of his participants, typically a young cat. The puzzle box is a kind of cage constructed so the cat can open the door from the inside by learning to work a manipulandum. The cat then escaped and ate some fish before being placed back in the box.²

桑代克想探究参与者是如何学习正确反应的。他描述了实验过程：在一个盒子里，猫必须拉动绳子末端的环或按钮：

Thorndike wanted to discover how the participant learns the correct response. He described what happens in a box in which the cat must pull a loop or button on the end of the string:

猫咪在盒子里到处乱抓，很可能会抓住绳子、绳圈或纽扣来打开门。渐渐地，所有其他不成功的冲动都会被消除，而导致成功行为的特定冲动则会被由此产生的快感所强化，直到经过多次尝试，猫咪被放进盒子里后，会立即抓住纽扣或绳圈。

Thorndike，1911 年，第 36 页

The cat that is clawing all over the box will probably claw the string or loop or button so as to open the door. And gradually all the other nonsuccessful impulses will be stamped out and the particular impulse leading to the successful act will be stamped in by the resulting pleasure, until, after many trials, the cat will, when put in the box, immediately claw the button or loop.

Thorndike, 1911, p. 36

桑代克将他的研究视为联想形成的研究之一：

Thorndike conceived his study as one of association formation:

然后，猫开始凭借其储存的本能冲动，找到成功的动作，并逐渐将其与盒子内部的感觉印象联系起来，直到完美地联系起来，以便它一遇到感觉印象就会做出动作。

Thorndike，1911年，第38页

Starting, then, with its store of instinctive impulses, the cat hits upon the successful movement, and gradually associates it with the sense-impression of the interior of the box until the connection is perfect, so that it performs the act as soon as confronted with the sense-impression.

Thorndike, 1911, p. 38

“反复尝试”，或者更确切地说是“尝试成功”——学习——恰如其分地描述了这些动物的行为。它们被放入培养皿中，尝试了各种熟悉的行为。猫可能会尝试挤过栅栏，抓挠笼子，或者把爪子伸进栅栏之间。最终，一只猫会用爪子抓住绳圈，然后拉动它，并发现自己的努力得到了回报。随着这些事件的重复，无用的行为逐渐消失，或者说消失，正确的行为在进入笼子后不久就会出现。

The phrase trial-and-error, or perhaps more exactly trial-and-success, learning aptly describes what these animals did. Placed inside, they tried out a variety of familiar behaviors. Cats were likely to try squeezing through the bars, clawing at the cage, or sticking their paws between the bars. Eventually, a cat pawed at the loop of string and so pulled on it, finding its efforts rewarded. As these events are repeated, useless behaviors die away, or extinguish, and the correct behavior is done soon after entering the cage.

巴甫洛夫学习

Pavlovian Learning

大约在同一时期，俄罗斯生理学家伊万·彼得罗维奇·巴甫洛夫（1849-1936）发现了巴甫洛夫条件反射，他此前已因其在唾液腺方面的研究而获得诺贝尔奖（1904 年）。在消化研究中，巴甫洛夫注意到一个奇怪的现象。在被用作参与者后，狗不仅会在食物出现时分泌唾液，还会对先前的中性刺激分泌唾液，例如看到喂食者的实验者。巴甫洛夫发现，不仅存在先天的刺激反应反射（将食物放入狗嘴中，狗会分泌唾液），还存在后天习得的反射（看到实验者时分泌唾液）。巴甫洛夫最初将这些称为心理反射（与生物反射相对）。巴甫洛夫并非心理学家，而是生理学家，他将学习研究视为研究大脑而非心智的窗口（Pavlov，1957 年）^。3

About the same time, Russian physiologist Ivan Petrovich Pavlov (1849–1936), who already had a Nobel Prize (1904) for his research on the salivary gland, discovered Pavlovian conditioning. During his digestion research, Pavlov noticed a curious phenomenon. After being used as a participant, a dog would not only salivate when food was presented but would also salivate to previously neutral stimuli—the sight of the experimenter who fed it, for example. Pavlov saw that not only are there innate stimulus-response reflexes (put food in a dog’s mouth and it salivates), but there are also learned reflexes (salivating at the sight of the experimenter). Pavlov originally called these psychical (as opposed to biological) reflexes. Pavlov was not a psychologist but a physiologist, and he viewed his studies of learning as windows on the brain, not the mind (Pavlov, 1957).³

巴甫洛夫将联想理论运用到他的实验环境中。他把“心理分泌”视为习得的联想；食物本身就会引起唾液分泌，并与实验者联系起来，实验者随后又引发唾液分泌。作为一名实验生理学家，巴甫洛夫深谙如何严密控制偶然发现，从而创建了一套严谨的实验范式。俄国革命期间，他勤奋工作。革命胜利后，新的苏维埃政权为巴甫洛夫建造了一个特殊的实验室，^被称为“寂静之塔”。为了排除不受控制的刺激，实验室配备了厚厚的玻璃窗、专门隔开的房间、嵌入沙子的地基以及一条充满稻草的护城河。实验期间，一只狗被拴在挽具上，实验者可以从另一个房间控制刺激呈现，并计算狗口中产生的唾液滴数（图4.2）。巴甫洛夫的实验装置可以完全控制动物的环境，并对学习情况进行客观的量化测量，他辉煌的研究计划由此开启。

Pavlov harnessed associationism to his experimental setting. Pavlov viewed “psychical secretions” as learned associations; the food, which already elicited salivation, became associated with the experimenter, who then elicited salivation. As an experimental physiologist, Pavlov knew how to put careful controls on his happenstance discovery, creating a rigorous experimental paradigm. He worked diligently through the Russian Revolution, and, after its victory, the new Soviet regime built Pavlov a special laboratory,⁴ known as the Towers of Silence, which, to ensure exclusion of uncontrolled stimuli, had thick glass windows, specially separated rooms, a foundation embedded in sand, and a straw-filled moat. During experiments, a dog was held in a harness, and from another room the experimenter could control stimulus presentations and count the drops of saliva produced in the dog’s mouth (Figure 4.2). Pavlov’s experimental setup allowed complete control of the animal’s environment and produced an objective quantitative measure of learning, and his brilliant research program began.

一些巴甫洛夫术语

能够引发生物反射的刺激（将食物放入口中、向眼睛吹一口气）是无条件刺激(US)。因此，生物引发的反射是无条件反应(UR)。通过与 US 配对而引发通常与 UR 几乎相同的反应的刺激称为条件刺激(CS)，因为它引发 UR 的能力取决于与 US 的配对。由 CS 引发的反射是条件反应(CR)。最后，US 通常被称为强化物，因为它与 CS 的配对会加强（强化）CS 引发 CR 的能力。

A stimulus that biologically elicits a reflex (food placed in the mouth, a puff of air at the eye) is an unconditional stimulus (US). The biologically elicited reflex is, therefore, an unconditional response (UR). A stimulus that through pairing with a US comes to elicit a response usually almost identical with the UR is called a conditional stimulus (CS) for its ability to elicit the UR is conditional on pairing with the US. The reflex elicited by a CS is a conditional response (CR). Finally, the US is often called a reinforcer, since its pairing with the CS strengthens (reinforces) the power of the CS to elicit the CR.

巴甫洛夫在他的实验中系统地研究了古老的哲学联想定律（图 4.3）。

Pavlov systematically worked through the old philosophical laws of association in his experiments (Figure 4.3).

以下是他的大部分基本发现，尽管它们绝不是巴甫洛夫研究的全部范围（Pavlov，1927；Todes，2014）。

Here are most his basic findings, though they by no means exhaust the range of Pavlov’s investigations (Pavlov, 1927; Todes, 2014).

条件反射的时间模式

休谟的邻近定律认为，如果两个想法同时出现，它们就会产生关联。在巴甫洛夫的实验范式中，我们可以通过控制呈现认知（CS）和感知（US）的时机来研究邻近定律。巴甫洛夫使用了两种认知（US）——食物（在消化过程中引起唾液分泌）和弱酸（在防御反射过程中引起唾液分泌）。他使用了许多认知（CS）——节拍器的咔哒声、音乐的音调、气味、沸腾的水声——尽管他没有使用著名的铃铛。我们在此假设 US 是食物，而 CS 是音符或音调。图 4.4显示了 US 和 CS 之间各种可能的时间关系。

Hume’s law of contiguity says that two ideas get associated if they occur at the same time. Within Pavlov’s experimental paradigm, the law of contiguity may be investigated by controlling when we present CS and US in relation to each other. Pavlov employed two USs— food, which elicited salivation as part of digestion, and mild acid, which elicited salivation as a defensive reflex. He used many CSs—the clicking of a metronome, musical tones, odors, the sound of bubbling water—though he did not use the famous bell. We will here assume the US is food and the CS is a musical note or tone. Figure 4.4 shows the various possible temporal relationships between the US and the CS.

根据邻近定律，我们可能认为同时呈现 US 和 CS 会最有效。然而，这样的配对并没有产生最快的条件反射。逆向条件反射（US 在 CS 开始之前呈现和移除）很难建立。相反，最佳条件反射发生在 CS 略早于 US 开始并与之重叠时。如果 CS 开始和结束都早于 US，则称为踪迹条件反射，这也很难建立。CS 和 US 必须重叠才能产生 CR 这一事实支持了邻近定律，但 CS 的开始必须先于 US 这一发现表明，CS 只有作为向生物体发出的信号时才会有效（Pavlov，1927）。

From the law of contiguity, we might expect simultaneous presentation of US and CS to be most effective. However, such pairings did not produce the fastest conditioning. Backward conditioning, in which the US is presented and removed before the CS begins, is hard to establish. Instead, best conditioning occurs when the CS begins slightly before the US and overlaps with it. If the CS begins and ends before the US, it is called trace conditioning, which is also difficult to establish. The fact that CS and US must overlap to create a CR supports the law of contiguity, but the finding that the onset of the CS must precede the US suggests that CSs become effective only if they act as signals to the organism (Pavlov, 1927).

复合调理

假设 CS 在 US 出现之前就开始了，也许早 30 秒，并且与 US 开始重叠。这被称为延迟条件反射；它类似于标准程序，只是 CS 开始的时间远早于 US。在这种情况下，CR 会建立，但会随着 CS-US 配对的进行而发生变化。起初，狗在 CS 开始时就开始流口水，但随着试验的继续，流口水直到 US 预计到达之前才开始。这一发现向巴甫洛夫 (1927) 表明时间本身可以充当 CS。引起流口水 CR 的控制 CS 不再仅仅是音调，而是音调加上一定时间的复合刺激。当两个刺激以这种方式控制反应时，这被称为复合条件反射。

Suppose the CS starts well before the US is presented, perhaps 30 seconds before, and overlaps US onset. This is called delay conditioning; it is like the standard procedure except that the CS begins far before the US. In this case a CR gets established, but it changes as CS-US pairings proceed. At first, the dog begins to salivate as soon as the CS begins, but as trials continue, the salivation does not commence until a little before the US is scheduled to arrive. This finding suggested to Pavlov (1927) that time itself could act as a CS. The controlling CS eliciting the salivation CR was no longer just the tone, but it was the compound stimulus of tone plus a certain passage of time. When two stimuli control a response in this way it is called compound conditioning.

巴甫洛夫发现，人们可以同时对两种或多种CS进行条件反射，例如，将光和声音与食物配对。另一个联想定律是生动性：一个给定的想法最容易与同时出现的生动想法联系起来，而不是与较弱的想法联系起来。巴甫洛夫的研究证实了这一点。如果两个CS的强度大致相等，并且单独呈现，经过条件反射后，每个CS都会引发CR。但如果其中一个比另一个强，例如明亮的光和昏暗的声音，那么只有更生动的刺激才会成为CS。当一种化合物的一个CS比其他CS强得多时，这种现象被称为遮蔽。

Pavlov found one could condition two or more CSs at the same time, for example, pairing a light and a tone with food. Another law of association was vividness: A given idea will get mostly associated with concurrent vivid ideas, rather than with weak ones. Pavlov’s work bore this out. If the two CSs were about equal in intensity, and if they were presented alone, after conditioning, each would elicit the CR. But if one were stronger than the other, a bright light and a dim tone, for instance, only the more vivid stimulus became a CS. This phenomenon, when one CS of a compound is much more powerful than the others, is called overshadowing.

兴奋和抑制

巴甫洛夫（1927）将对特定正向反应（例如唾液分泌）的条件反射称为兴奋性条件反射。US（刺激）自然会刺激 UR（刺激上皮层），经过几次配对后，CS（刺激后皮层）也会兴奋。巴甫洛夫还认为，CR（刺激后皮层）的活动既可以被抑制，也可以被激发。

The conditioning of a distinct positive response, such as salivation, Pavlov (1927) called conditioning of excitation. The US naturally excites the UR, and, after several pairings, so does the CS. Pavlov also believed that the action of CRs could be inhibited as well as excited.

回到延迟条件反射，我们看到了抑制的作用。在多次 CS-US 配对之后，CR 仅发生在 CS 间隔接近尾声时，即 US 的预期时间。那么，为什么 CS 开始前的早期 CR 会消失呢？音调的第一部分是否只是断开连接，与 CR 分离了？巴甫洛夫认为不是。他认为，音调的早期部分主动抑制了 CR 的发生。作为证据，巴甫洛夫指出了脱抑制：如果在 CS 的早期阶段，在任何 CR 发生之前，引入一个全新的刺激，例如光的闪光，CR 就会发生。现在，如果 CR 只是与早期 CS 断开连接，光不会引起唾液分泌，就像在实验室外一样。但是，在 CS 期间光闪烁时意外出现唾液分泌，这向巴甫洛夫表明，光干扰了皮质的抑制过程，释放了被抑制的对音调的 CR。

Returning to delay conditioning, we see inhibition in action. After many CS-US pairings, the CR occurred only near the end of the CS interval, at the anticipated time of the US. Why then did the earlier CRs to the start of the CS disappear? Was the first part of the tone simply disconnected, disassociated from the CR? Pavlov thought not. He believed that the early part of the tone actively inhibited the CR from taking place. As evidence, Pavlov pointed to disinhibition: If, during the early stages of the CS, before any CR would occur, a wholly new stimulus is introduced, such as a light flash, the CR would occur. Now if the CR had simply been disconnected from the early CS, the light would not cause salivation, any more than it would outside the laboratory. But the unexpected appearance of salivation when the light flashed during the CS indicated to Pavlov that the light interfered with the cortical process of inhibition, releasing the suppressed CR to the tone.

最重要的抑制形式是消退。如果我们只呈现一个既定的CS，而不进行进一步的强化，会发生什么？CR会持续一段时间，然后慢慢减弱，最终降至零。在这里，似乎可以这样说：CS已经与 CR 分离，甚至 CS 不再是 CS 而是中性刺激——反射实际上已经消失。

The most important form of inhibition is extinction. What happens if we present an established CS alone, without further reinforcement? The CR occurs for a while, slowly weakens, and finally falls to zero. It seems plausible to say here that the CS has been disassociated from the CR, even that the CS is no longer a CS but a neutral stimulus—the reflex has literally extinguished.

但进一步的研究表明，CS-CR关联仍然存在，只是被消退程序抑制了。如果我们只是将动物放置一段时间，然后再次呈现CS，唾液分泌就会重新出现。条件反射已经自发恢复。这表明该反射并非因消退而习得，而是被主动抑制了；这也表明抑制是暂时的。

But further investigation reveals that the CS-CR association is still present, only inhibited by the extinction procedure. If we simply put the animal aside for some time and present the CS again, salivation reappears. The conditional reflex has spontaneously recovered. This indicates that the reflex was not unlearned by extinction but was actively inhibited; it also indicates that inhibition is temporary.

另一种证据支持巴甫洛夫的论点。假设我们在唾液分泌量为零之后继续进行消退，并反复呈现条件反射。如果消退仅仅是对条件反射的遗忘，那么额外的非强化配对就不会产生任何效果：一旦条件反射变成了中性刺激，就无法再说它变得更加中性了。相反，巴甫洛夫发现，消退持续的时间越长，条件反射自发恢复所需的时间就越长，这表明额外的非强化配对会加深条件反射，从而延长对条件反射的抑制。

Another kind of evidence supports Pavlov’s argument. Suppose we continue extinction well beyond the point at which zero salivation occurs, presenting the CS over and over. If extinction were simply unlearning the conditional reflex, additional unreinforced pairings would have no effect: Once a CS has become a neutral stimulus, it makes no sense to say it has become more neutral still. Instead, Pavlov found that the longer extinction continued past zero, the longer it took for the CR to spontaneously recover, indicating that additional unreinforced pairings deepened and so prolonged the inhibition of the CR.

概括与歧视

休谟的另一条联想定律是相似性：相似的想法很容易联系起来。巴甫洛夫发现 CR 不仅发生在训练过的 CS 上，也发生在类似的刺激上。假设我们训练一条狗对每秒 1,000 个周期的音调分泌唾液，然后呈现接近或远离 1,000 个周期的音调（或高或低）。我们发现 CR 发生在相关音调上，并且新音调越接近原始 CS，CR 越大；越远，CR 越小。CR 扩展到像 CS 这样的刺激称为泛化，而随着测试刺激与 CS 越来越不同，CR 的逐渐减弱称为泛化梯度。

Hume’s other law of association is similarity: Similar ideas are easily associated. Pavlov found that CRs occur not just to the trained CS, but to similar stimuli, too. Suppose we have trained a dog to salivate to a tone of 1,000 cycles per second, and then present tones near or far from 1,000 cycles, either above or below. We find that the CR occurs to the related tones, and that the closer the new tone is to the original CS, the greater the CR; the farther away, the less the CR. The extension of a CR to stimuli like the CS is called generalization, and the gradual weakening of the CR as the test stimuli increasingly differ from the CS is called the generalization gradient.

我们可以将兴奋和抑制相互抵消，从而产生刺激辨别。假设我们调节唾液分泌，使其适应屏幕上呈现的发光圆圈的刺激刺激（CS+）（Pavlov，1927）。如果我们随后呈现从近圆形到极椭圆形的椭圆，就会发现一个泛化梯度。然而，在呈现 CS+ 之后，我们立即反复呈现一个极椭圆形，并且从不强化它（CS-）。抑制会逐渐增强，直到椭圆形（CS-）。这样，我们就创造了刺激辨别。动物对 CS+ 有反应，而对 CS- 没有反应。

We can set excitation and inhibition against one another to produce stimulus discrimination. Suppose we condition salivation to a CS (CS+) of a luminous circle presented on a screen (Pavlov, 1927). If we then present ellipses varying from near-circular to extremely elliptical, we find a generalization gradient. However, right after presenting the CS+, we repeatedly present an extreme ellipse alone and never reinforce it (CS-). Inhibition will build up to the ellipse (CS-). We have created stimulus discrimination. The animal responds to CS+ and does not respond to CS-.

了解了泛化之后，我们现在可以提出以下问题：如果我们继续进行辨别过程，同时使连续的CS-越来越接近圆形，会发生什么？兴奋和抑制的泛化梯度将开始发生冲突：CS-像圆形，会在一定程度上引发唾液分泌，而作为椭圆形则会引发抑制；CS+类似于椭圆，会在一定程度上引发抑制，而作为圆形则会引发唾液分泌。随着CS-越来越像CS+，这两种对抗趋势将会发生冲突。如果进一步深入，巴甫洛夫的狗“会呈现所有”“它会尖叫、扭动、撕扯仪器并狂吠不止（Pavlov，1927，第291页）。”巴甫洛夫预见了行为疗法，他提出，由于精神病理学是通过条件反射习得的，因此它应该可以通过重新条件反射“消除”。

Knowing about generalization, we can now ask the following question: What happens if we continue the discrimination procedure while making successive CS-s more and more circular? The generalization gradients of excitation and inhibition will begin to clash: CS-, being like a circle, will to some degree elicit salivation, while as an ellipse it will elicit inhibition; CS+, being similar to an ellipse, will to some degree elicit inhibition, while as a circle it will elicit salivation. As CS- increasingly resembles CS+, the antagonistic tendencies will clash. Pushed far enough, Pavlov’s dog “presented all the symptoms of acute neurosis.” It squealed, it wiggled, it tore at the apparatus and barked violently (Pavlov, 1927, p. 291). Pavlov anticipated behavior therapy by proposing that because psychopathology is learned by conditioning it should be “undoable” by reconditioning.

阐明学习

Elucidating Learning

案例研究：大鼠的洞察力

Case Study: Insight in Rats

托尔曼（Tolman）的一项实验（1932；托尔曼和洪齐克（Honzik）的实验，1930）采用了黄金时代研究学习的主流方法——迷宫。实验的问题是，在学习了迷宫之后，当实验者改变路线时，老鼠是否能够运用它们的知识找到通往目标的最短路线。更深层次的问题在于，如何将托尔曼的詹姆斯式观点与赫尔的数学版本进行对比，前者认为生物体能够智能地适应环境，后者认为动物是连接刺激和反应的机械连接器。

One experiment (Tolman, 1932; Tolman & Honzik, 1930) used the dominant methodology for studying learning during the Golden Age, the maze. At issue was whether having learned a maze, rats could use their knowledge to figure out the shortest route to their goal when the experimenter altered the routes. The deeper issue was contrasting Tolman’s Jamesian view of organisms as adapting intelligently to their environments versus Hull’s mathematical version of Thorndike’s view of animals as mechanical connectors of stimuli to responses.

桑代克曾用他的效果定律提出了一种对目标导向行为的机械解释，该定律“印证了”某些刺激→反应之间的联系，赫尔对其进行了改进。常识会认为桑代克的一只猫找到了获得自由和获得所需食物的方法，而刺激→反应学习理论则认为，获取食物的快感只是将谜盒内的刺激与拉动盒顶悬挂的绳子的行为粘合在一起，这样拉绳子就成为了一种自动反应，会自动消失。赫尔说，实际上，桑代克的学习与巴甫洛夫的学习没有什么不同。巴甫洛夫的狗并不是为了获取食物而流口水，因为流口水并不能使食物变得可得。同样，桑代克的猫也不是为了逃脱和获取食物而拉绳子，它们只是重复了受谜盒中刺激控制的动作。赫尔 (1937) 表示，根本不存在有目的的行为。

Thorndike had proposed a mechanistic explanation of goal-directed behavior with his Law of Effect that “stamped in” certain Stimulus → Response connections, and Hull refined it. Where common sense would say that one of Thorndike’s cats had figured out a way to freedom and a desired bit of food, S → R learning theory said that the pleasure of getting food had simply glued together the stimuli inside the puzzle box to the pulling of the string dangling from the box’s top, so that the string-pulling became an automatic response that simply ran itself off. In effect, said Hull, Thorndikian learning was no different from Pavlovian learning. Pavlov’s dogs did not salivate in order to get food, because salivating did not cause the food to become available. Similarly, Thorndike’s cats did not pull the string in order to escape and get food, they simply repeated a movement that had come under control of the stimuli in the puzzle box. There was no such thing as purposive behavior, said Hull (1937).

托尔曼赞同詹姆斯关于意识是“目的的斗士”的观点，但他不同意这种观点，他认为目的是生命的基本事实，而非需要解释或消除的谜团（托尔曼，1925、1932、1959）。他遵循感知实在论，进一步指出目的并非我们从行为中推断出的神秘内在心理状态，而是行为中可观察到的特征。托尔曼的许多实验，包括“老鼠的‘顿悟’实验”，都是为了证实这些论断。

Tolman, agreeing with James that consciousness is a “fighter for ends,” disagreed, argued that purpose is a basic fact of life, not a mystery in need of explanation or elimination (Tolman, 1925, 1932, 1959). Following perceptual realism, he further claimed that purpose is not a mysterious inner state of mind that we infer from behavior, but an observable feature of behavior. Many of Tolman’s experiments, including “‘Insight’ in rats,” were designed to establish these claims.

图 4.5展示了迷宫。迷宫路径是一块狭窄的木板，老鼠沿着它从起始平台跑到目标平台，在那里它们可以找到食物。其中两个迷宫的木板有侧面，所以老鼠看不到它们所在的房间；第三个迷宫的木板没有墙。

Figure 4.5 shows the maze. The paths were narrow boards along which the rats ran from the starting platform to the goal platform where they found food. In two of the mazes, the boards had sides, so the rats could not see the room they were in; in the third, there were no alley walls.

首先，受试者被迫跑完所有三条路径。决定性时刻到来时，实验人员在路径2的返回点与路径1的返回点以及路径3的返回点与路径1的返回点之间设置了一个障碍物。遇到这个障碍物时，老鼠会回到选择点。它会选择哪条路线？（图4.6）

First, the participants were forced to run all three paths. The decisive moment came when the experimenters placed a barrier halfway between where the return of Path 2 met Path 1 and the return of Path 3 met Path 1. Meeting this barrier, a rat would return to the choice point. What route would it choose? (Figure 4.6).

Hull 表示，老鼠已经学会了一种习惯家族层级结构(Hull, 1934)，即一组由控制刺激以不同强度引发的反应。在这种情况下，每个 R 实际上都是老鼠沿着小巷一步一步移动时产生的一系列反应。在这个层级结构中，反应的整体强度由反应链的长度决定，因此走路径 1 的强度最强，走路径 3 的强度最弱，走路径 2 的强度居中。因此，当老鼠在路径 1 的末端附近遇到障碍物时，它会退回到选择点，该点的刺激会引发下一个最强的反应链，老鼠会走路径 2，再次遇到障碍物，再次退回到选择点，这将触发走路径 3，最后参与者会到达目标箱并进食。

Hull said that the rat had learned a habit family hierarchy (Hull, 1934), that is, a set of responses provoked with varying strength by a controlling stimulus. In this case, each R is actually a chain of responses as the rat moves paw step by paw step down an alley. In this hierarchy, overall strength of response was determined by the length of the response chain, so that taking Path 1 was strongest and taking Path 3 was the weakest, with taking Path 2 having a middling value. Therefore, when the rat met the barrier near the end of Path 1, it would back up to the choice point, whose stimuli would evoke the next strongest response chain, and the rat would take Path 2, meet the barrier again, retreat again to the choice point, which would trigger taking Path 3, and finally the participant would reach the goal box and eat.

另一方面，托尔曼认为，动物并非反射性机器；相反，它们始终是目标导向的斗士，为最终目标而战。因此，在迷宫中，在某个地点找到食物，使该地点成为需要记住和追寻的目标。生物体学习迷宫的空间布局——一张认知地图（托尔曼，1948），而不仅仅是一组有序的反射。拥有了这张学习地图后，老鼠会展现出洞察力，能够看到障碍物阻挡了通过路径1和路径2到达目标的通道，并立即选择路径3。这种洞察力在托尔曼和洪齐克的第三个迷宫中被发现。

Tolman, on the other hand, believed that animals are not reflexive machines; rather, they are always goal-directed fighters for ends. Thus, in a maze, finding food in a certain spot made that spot a goal to be remembered and sought. Organisms learn the spatial layout—a cognitive map (Tolman, 1948)—of the maze, not simply a set of ordered reflexes. In possession of its learned map, a rat would show insight, able to see that the barrier blocked access to the goal via both Path 1 and Path 2 and would right away take Path 3. Such insight was found in the third of Tolman’s and Honzik’s mazes.

空间学习的进一步证据来自潜在学习的研究（Thistlethwaite，1951）。在这些实验中，老鼠穿过没有食物作为诱饵的迷宫，因此没有强化S→R链。然而，当食物被放置在迷宫中的某个位置时，经验丰富的老鼠可以立即找到食物，这表明它们在没有奖励的情况下获得了认知地图（Tolman，1948）。在人类中，Tolman、Hall和Bretnall（1932）表明，如果疼痛在迷宫中标记了通往目标的道路，它就可以像奖励一样发挥作用。Tolman几乎区分了事件的信息价值（告诉生物体它是否在做正确的事情）和情感价值（它的感受）。

Further evidence for spatial learning came in studies of latent learning (Thistlethwaite, 1951). In these experiments, rats ran through mazes that were not baited with food, so that no S → R chains were reinforced. Nevertheless, when food was placed in a location in the maze, experienced rats could immediately find their ways to it, showing that they had acquired a cognitive map in the absence of reward (Tolman, 1948). In humans, Tolman, Hall, and Bretnall (1932) showed that pain could operate like a reward if it marked the way toward a goal in a maze. Tolman came close to distinguishing the information value of an event—telling the organism whether it’s doing the right thing—from its affective value—how it feels.

“大鼠的‘洞察力’”的最后一个方面指向学习的神经科学。回想一下，实验中使用了三个迷宫。其中两个迷宫的爪子路径是迷宫被围起来，所以参与者看不到迷宫外面，也没有表现出顿悟的迹象，但行为符合 S → R 理论。只有在第三个迷宫中才表现出顿悟，老鼠可以看到迷宫所在的实验室。如果没有迷宫外的线索，前两个迷宫就无法绘制出有用的地图，所以老鼠必然会回忆起它们到达目标的步骤。有了线索，参与者就能绘制地图，并能进行顿悟。老鼠在两种情况下都成功找到了食物，这表明学习肯定不止一种，至少有 S → R 学习和空间位置学习。现代认知神经科学的一项里程碑式发现⁵表明，皮层下的一种叫做海马体的结构负责制作认知地图（O'Keefe & Nadel，1978，1979）。

A final aspect of “‘Insight’ in rats” points to neuroscience of learning. Recall that there were three mazes used in the experiment. In two of them, the paw paths were walled, so the participants could not see outside the maze, and they showed no signs of insight, but behaved in accord with S → R theory. Insight was shown only in the third maze, in which the rats could see the laboratory in which the maze stood. Without the extra-maze cues, in the first two mazes no useful map could be built up and so the rats were necessarily thrown back on recalling their steps to the goal. With the cues, the participants constructed maps and could consult them insightfully. That the rats succeeded under both conditions in finding the food shows that there must be more than one kind of learning, at least S → R learning and spatial location learning. One of the landmark findings in modern cognitive neuroscience⁵ showed that a subcortical structure called the hippocampus is responsible for making cognitive maps (O’Keefe & Nadel, 1978, 1979).

行为的实验分析

Experimental Analysis of Behavior

斯金纳坚持认为实验控制是检验科学真理性的终极标准，体现在两个方面。首先，能够调整自变量并观察到行为的变化，表明你的理论主张是正确的。其次，斯金纳始终认为科学应该为社会带来实际回报，即控制人们的行为以造福人类（Skinner，1971，1986）。他批评认知科学未能做到这一点（Skinner，1985）。“分析”一词使他的研究体系秉承了将复杂事物和行为分解为各个组成部分的传统，例如物理学和化学。这使他反对将复杂现象视为其各部分之和以外的其他传统，例如格式塔心理学。“行为”一词宣誓效忠激进行为主义：不涉及有机体的内部状态。

Skinner maintained that experimental control was the ultimate test of truth in science in two ways. First, being able to tweak independent variables and see changes in behavior showed that your theoretical claims were correct. Second, Skinner always believed that science should pay off in practical returns to society, i.e., controlling peoples’ behavior for their own good (Skinner, 1971, 1986). He criticized cognitive science for not being able to do so (Skinner, 1985). The word “analysis” sets his program in the tradition of breaking complex things and behaviors into their component parts, as in physics and chemistry. It set him against traditions that see complex phenomena as more than the sum of their parts, as in Gestalt psychology. The word “behavior” pledges allegiance to radical behaviorism: There will be no reference to inner states of the organism.

斯金纳批判现有的心理学，并试图用自己的观点取而代之。他在研究生院发明的方法论——斯金纳箱（他称之为实验空间）——很快取代了大多数其他研究动物行为的方法。他对现有方法的批判至今仍具有重要意义，因为他指出了实验方法一个重要但经常被忽视的特征：你的方法的特征看似直觉的东西，实际上可能比你想象的更能控制参与者的行为。我称之为方法陷阱；格式塔心理学家也有类似的担忧。

Skinner was critical of existing psychology and sought to replace it with his own ideas. The methodology he invented in grad school, the Skinner box (he called it an experimental space), soon replaced most other way of investigating animal behavior. His critique of existing methods remains important because he pointed out an important but often overlooked feature of experimental methods, that features of your method that seem intuitive to you may control more of your participant’s behavior than you think. I call this the trap of the method; Gestalt psychologists had similar concerns.

斯金纳最初的担忧与生物学有关，主要集中在巴甫洛夫条件反射上。巴甫洛夫研究了一种固定的反射——唾液分泌，它属于一组预先存在的、进化而来的功能，这些功能专注于消化这一单一功能。斯金纳担心，巴甫洛夫的条件反射定律可能只适用于唾液分泌反射，而无法提供适用于所有物种的学习原则。斯金纳的担忧并非空穴来风，因为物种在学习方面存在差异是很常见的。例如，托尔曼和赫尔研究了老鼠学习迷宫，但它们天生就适应学习迷宫，因为它们生活在分支隧道中，因此必然擅长于穿越这些隧道。

Skinner’s first worry concerned biology and focused on Pavlovian conditioning. Pavlov studied a fixed reflex, salivation, that was part of a preexisting evolved set of functions devoted to one function, digestion. Skinner feared it was possible that Pavlov’s laws of conditioning might be true only for the salivary reflex and would not provide learning principles valid across species. Skinner’s worry was well placed, as species differences in learning are common. For example, Tolman and Hull studied rats learning mazes, but learning mazes is something they are preadapted for, because they live in, and thus must be good at, navigating branching tunnels.

另一个担忧与定义行为有关。在本例中，斯金纳的目标主要在于迷宫学习，他提出了一些在其他方面被忽视的担忧。迷宫学习中最常用的因变量是每次运行的错误次数，就像桑代克的试错学习一样。问题在于，错误是我们强加于动物的人类道德观念。什么算作“错误”？在某个特定的选择点，老鼠可能会完全走进死胡同。或者它可能会走进死胡同一小段距离。或者它可能会走进死胡同但又折了回来。或者它可能会先在入口处停顿一下再继续前进。或者它可能会放慢速度，看看小巷，但继续前进。这些都是错误吗？还是其中一些？我们如何知道？通常的答案是操作性定义，将错误定义为，例如，老鼠将50%或更多的身体（不包括尾巴）走进死胡同。但这是心理学家对自然行为而非科学观察的武断判断。最好从实验本身中消除“错误”，并排除人为判断。

Another concern had to do with defining behaviors. In this case, Skinner’s target was primarily maze learning, and he raised otherwise ignored concerns. The most frequently used dependent variable in maze learning was the number of errors made on each run, as in Thorndike’s trial and error learning. The problem is that error is a human moral concept that we impose on the animal. What counts as an “error?” At a given choice point a rat might go all the way down a blind alley. Or it might go some partial distance down the blind alley. Or it might start down the blind alley but back out. Or it might pause at the entrance before continuing. Or it might slow down, look down the alley, but keep going. Are all of these errors? Some of them? How could we know? The usual answer was operational definition, defining an error as, for example, the rat getting 50% or more of its body (not counting the tail) down a blind alley. But this is an arbitrary judgment a psychologist passes on natural behavior, not a scientific observation. It would be better to eliminate “error” from the experiment itself and leave human judgment out.

其他方法则显得人为，因为它们将连续的行为流程切割成任意的试验，从而打断了连续的行为流。我们选择何时呈现US和CS；我们决定何时完成老鼠的迷宫测试并将其放回起始盒；我们打断猫的进食并将其放回谜题盒。这样一来，我们的数据就变成了一系列离散数字，用于测量试验中的唾液滴、运行中的错误或谜题盒中的解答时间。最好在整个实验过程中对某个行为进行连续记录。

Other methods were artificial in that they interrupted the continuous flow of behavior by chopping it up into arbitrary trials. We choose when to present US and CS; we decide when a rat’s maze run is done and drop it back in the start box; we interrupt the cat’s eating and put it back in the puzzle box. Our data are then a series of discrete numbers measuring saliva drops on a trial, errors on a run, or time-to-solution in a puzzle box. It would be better to have a continuous record of a behavior during the whole course of an experiment.

综合这些要求，斯金纳希望研究不属于既存生物系统的行为学习，并在一个行为被持续记录而非评判的环境中开展研究。经过大量的反复试验（Skinner, 1959; Lattal, 2004），他提出了斯金纳箱。

Putting these requirements together, Skinner wanted to study the learning of behavior not part of a preexisting biological system, in a setting in which that behavior was recorded, not judged, continuously. After a lot of trial and error on his part (Skinner, 1959; Lattal, 2004), he came up with the Skinner box.

斯金纳箱主要有两种，一种用于大鼠，一种用于鸽子（图4.7）。这两种箱通常都由透明塑料制成，有些箱底配有电击网格，箱体大小刚好能让实验对象舒适地躺下并四处走动。箱体一侧墙上有一个可以向喂食杯中分配食物或水的弹匣、一个扬声器、一盏或多盏用于辨别刺激的灯，以及动物必须操作才能获取食物或水的操作器。用于大鼠的操作器是一个水平伸出的杠杆对鸽子来说，它是一个嵌在墙上的发光圆盘或钥匙。操作操纵器可以激活食物或水匣。对于研究人员来说，斯金纳箱的一个额外好处是无需监督。实验开始后，计算机可以记录数据并操纵刺激和奖励变量。

There are two main kinds of Skinner boxes, for rats and for pigeons (Figure 4.7). Both boxes are usually made of clear plastic, some with a shock grid on the floor, and are just large enough to hold the subject comfortably and allow it to move around. One wall contains a magazine that can dispense food or water to a feeding cup, a speaker, a light or lights for discriminative stimuli, and the manipulandum that the animal must operate to get food or water. The manipulandum for the rat is a horizontal lever sticking out of the wall; for pigeons it is a lighted disc or key set into the wall. Operation of the manipulandum activates the food or water magazine. For researchers, a side benefit of Skinner boxes is that they need no supervision. Once the experiment is underway, computers can record data and manipulate stimulus and reward variables.

除了一个例外，这些盒子都符合斯金纳的要求。动物一旦被放入盒子中，就可以自由活动，直到实验结束前都无需处理，因此无需进行单独试验。对于大鼠来说，按压杆是一种在野外不太可能发生的随意行为。对于鸽子来说，啄食发光的钥匙似乎是一种随意行为，但实际上并非如此，因为啄食是鸽子的进食方式。事实证明，啄钥匙行为中存在巴甫洛夫式的成分。即使啄食和食物递送之间没有后验关系，仅仅点亮钥匙并呈现食物也能引发啄食行为，这种现象被称为自我塑造(Rescorla, 1984)。

With one exception, these boxes met Skinner’s requirements. Once placed in the box, animals move about freely and need not be handled until an experiment is over, so that separate trials are not imposed. For the rat, bar-pressing is an arbitrary behavior unlikely to occur in the wild. In the case of the pigeon, pecking a lighted key appears to be an arbitrary behavior, but it really is not because pecking is how pigeons feed. It turns out that there is a Pavlovian component to key pecking. Simply lighting the key and presenting food recruits pecking even without a contingency between the peck and food delivery, a phenomenon known as autoshaping (Rescorla, 1984).

可以简单地将老鼠放入斯金纳箱并等待它压杆，但在实践中，实验者会逐渐将行为塑造成最终形式。塑造对于激进行为主义、行为分析的专业实践方面至关重要。实验者有一个操作食物漏斗的按钮。当老鼠探索盒子时，实验者在老鼠位于盒子饲料盒末端时将食物放入喂食盒中，然后仅当老鼠面向饲料盒时，然后仅当老鼠靠近杆时，然后将自己抬到杆上方时，老鼠自己很可能在此时第一次压杆。此时，可由计算机控制强化，而累积记录器（笔下展开的连续纸卷）会实时记录每次压杆的情况。

It’s possible to simply put a rat in a Skinner box and wait for it to press the bar, but in practice experimenters gradually shape the behavior into its terminal form. Shaping is vital to the professional practice side of radical behaviorism, behavior analysis. The experimenter has a button that operates the food hopper. As the rat explores the box, the experimenter drops food in the feeding magazine when the rat is in the magazine’s end of the box, then only when the rat is facing the magazine, then only when it gets closer to the bar, then raises itself above the bar, when the rat’s own first bar press will probably occur. At this point, reinforcements can be computer-controlled while the cumulative recorder—a continuous roll of paper unfurling under a pen—notes each bar press as it occurs in real time.

斯金纳箱也消除了迷宫中谜题盒残留的拟人化。桑代克认为，他的猫和他一样，把盒子看作一个需要逃脱的迷宫陷阱，因此，未能达到这一目的的行为是令人遗憾的错误。迷宫，同样，也提出了一个挑战，即老鼠像运动员一样“努力”缩短达到目标的时间。另一方面，在斯金纳箱中，参与者只是发现自己身处一个新环境，行为举止与平常无异，没有固定目标，因此不会犯任何“错误”。当杠铃推举后出现强化结果时，杠铃推举的频率会发生变化。理论上可以避免提及目标和努力（Skinner, 1984）。

The Skinner box also eliminates the residual anthropomorphism of puzzle boxes in mazes. Thorndike assumed that his cats saw the box as he did, a puzzling trap to escape, and thus that behaviors failing to accomplish this end were regrettable errors. Mazes, too, suggest a challenge as rats “strive” like athletes to decrease their time to goal. In the Skinner box, on the other hand, the participant simply finds itself in a new environment and behaves as it usually would with no fixed goal and thus no “errors” to commit. When the consequence of reinforcement follows a bar press, the frequency of bar pressing changes. Theoretical references to goals and striving can be avoided (Skinner, 1984).

斯金纳对学习争议的分析颇具创新性。与赫尔一样，他坚持不参考心理状态来定义行为，但与托尔曼一样，他认为反射层面或分子层面的定义并不适用。相反，他定义操作行为时，不是根据动作本身，而是根据其结果。在这种情况下，作者对语言的熟悉为他提供了很大的帮助。

Skinner’s analysis of learning controversies was innovative. Like Hull, he wanted to stick to behaviors defined without reference to mental states, but, like Tolman, he saw that reflex level, or molecular, definitions would not do. Instead, he defined operant behaviors not in terms of movements made, but in terms of what they accomplish. In this case, his writer’s familiarity with language served him well.

考虑以下行为：说出“乔治·华盛顿”、“美国第一任总统”、“约克镇战役的胜利者”；或者像我刚才那样打字；或者像以前上课那样把它们写在黑板上；或者举起一张一美元钞票，指着肖像说“他！”。这些行为涉及身体不同部位的不同肌肉运动，但显然应该将它们归为一类，因为按照传统说法，它们都指代乔治·华盛顿这个人。但斯金纳箱中的老鼠也存在类似情况。老鼠可能会用左爪按压横杆；用双爪按压；用头部侧面按压；如果绑住前爪，则用后爪按压。作为动作，这些行为各不相同，但它们是相同的，因为它们得到的是同一种东西。此外，老鼠可能会做出同样的动作来按压横杆获取食物和水，但它们不应被视为同一种反应，因为它们会受到不同后果的强化。

Consider the following behaviors: Saying “George Washington,” “The first president of the United States,” “The victor at Yorktown”; or typing these words as I just did; or writing them on a blackboard as I used to in class; or holding up a dollar bill and pointing at the portrait while saying “Him!” These behaviors involve different muscle movements of different parts of the body, but clearly they should be grouped together because they all refer, in traditional parlance, to the person George Washington. But something similar is true about a rat in a Skinner box, too. A rat might press the bar with its left paw; with both paws; with the side of its head; with its rear paw(s) should we tie up its front paws. As movements, these behaviors are different, yet they are the same because they get the same thing. In addition, a rat might make the same movement to bar press for food and to bar press for water, but they should not be counted as the same response because they are reinforced by different consequences.

斯金纳提出，任何给定的操作不应定义为一种运动或反射，而应定义为由相同辨别刺激控制并由相同结果强化的一类行为。因此，所有对华盛顿的提及尽管它们的运动地形不同，但都是同一操作，因为它们的控制刺激是相同的。老鼠为了食物而做出的所有运动地形不同的按杆动作都是同一操作，因为强化物是相同的。最后，为了食物和水而做出的运动地形相同的按杆动作属于两个不同的操作，因为它们的强化物不同。事实上，操作可以在没有任何运动的情况下发生。瘫痪的人可以学习使用连接到他们脑电波的计算机界面来操作他们环境中的设备，例如电视或炉灶（Birnbauer，2006；Ruf＆Birnbauer，2012；Willett，2021 ⁶）。

Skinner proposed that any given operant should be defined not as a movement or a reflex, but as a class of behaviors controlled by the same discriminative stimuli and reinforced by the same consequence. Thus, all the references to Washington are the same operant despite their different movement topographies because their controlling stimulus is the same. All the topographically distinct bar presses of the rat for food are the same operant because the reinforcer is the same. Finally, the topographically identical bar presses for food and water belong to two different operants because their reinforcers are different. Indeed, an operant can occur without any movements at all. Paralyzed people can learn to use computer interfaces linked to their brain waves to operate devices in their environments, such as televisions or stovetops (Birnbauer, 2006; Ruf & Birnbauer, 2012; Willett, 2021⁶).

斯金纳方法论的最后一个方面涉及研究个体还是群体。心理学从一开始就以研究个体为特色，无论是早期实验心理学家的内省报告，还是临床心理学家的测试结果。其他新兴的社会科学，例如社会学心理学家通过收集和分析统计数据来研究群体。在理论方面，心理学家撰写了关于个体生物体实体和过程的文章，例如注意力、记忆力或学习能力。然而，随着时间的推移，尽管心理学继续对个体进行理论研究，但它收集了群体数据，并使用统计数据来比较群体结果。

A final aspect of Skinner’s methodology concerns whether to study individuals or groups. From its beginning, psychology distinguished itself by studying individual human beings, whether in introspective reports from the first experimental psychologists or the test results of clinical psychologists. Other nascent social sciences, such as sociology and economics, studied groups of people through collecting and analyzing statistics. On the theoretical side, psychologists wrote about entities and processes belonging to individual organisms, such as attention, memory, or learning. However, over time, although psychology continued to theorize about individuals, it collected group data and used statistics to compare group results.

然而，将众多参与者的行为汇总成群体平均值可能会掩盖个体行为的实际情况。例如，Voeks (1954) 研究了人类的眨眼条件反射，这通常被视为巴甫洛夫条件反射的一种形式。在这种装置中，参与者的头部被固定在一个类似眼科医生使用的框架中。US 指的是一股气流，引发参与者眨眼的 UR（重复性反应）。US 与 CS 配对，例如闪光或音调。最终，CR 会被习得。Voeks 问道，它是如何习得的，是逐渐习得的还是在一次试验中习得的？Hull 的理论认为，学习总是循序渐进的，习惯强度会随着强化次数的增加而增强。然而，其他理论认为，在习得联想的场合，联想的强度会在瞬间达到最大强度。当 Hullians 让受试者分组并绘制不同试验中 CR 的数量时，他们得到了平滑的学习曲线，该曲线逐渐增加，直到每个人都对 CS 眨眼。

However, there is a danger that aggregating the behavior of many participants into group averages may obscure what’s going on in individuals. For example, Voeks (1954) studied human eyeblink conditioning, usually treated as a form of Pavlovian conditioning. In this setup, the participant’s head is held in a framework like those used by eye doctors. The US is a puff of air that elicits a blink UR from the participant. The US is paired with a CS, such as a light flash or a tone. Eventually, a CR will be learned. Voeks asked how it’s learned, gradually or on a single trial? Hull’s theory says learning is always gradual, as habit strength builds up with number of reinforcements. However, other theories held that on occasions when associations were learned, the strength of association reached full strength in a single instant. When Hullians ran groups of subjects and plotted the number of CRs across trials, they got smooth learning curves that increased gradually until everybody was blinking to the CS.

Voeks 认为，这些学习曲线是将个体行为聚合到群体数据中造成的人为现象。在任何给定的试验中，参与者都有可能将 CS 和 US 联系起来，在试验前不给出 CR，试验后也始终给出 CR。换句话说，个体会表现出“全有或全无”的学习，而不是习得曲线。由于个体会在不同的试验中从未学习状态过渡到已学习状态，因此群体中 CR 的累积数量会缓慢上升，直到每个人都处于已学习状态。进行实验并查看每个人的数据非常简单，Voeks 发现她的分析是正确的。参与者从不给出 CR 过渡到在单次试验中每次都给出 CR。这符合 Brewer (1974) 后来提出的学习认知图景，即受试者对学习过程表现出洞察力，有意识地理解 CS 可以预测 UR，然后产生适当的行为 (Leahey, 1979)。

Voeks argued that these learning curves were artifacts caused by aggregating individual behaviors into group data. It was possible that on any given trial, there was some probability that a participant would associate CS and US, not giving the CR before that trial and never failing to give it after that trial. In other words, individuals would show all-or-nothing learning rather than a curve of acquisition. Because individuals would go from the unlearned state to the learned state on different trials, the cumulative number of CRs in the group would rise slowly until everyone was in the learned state. It was a simple matter to run an experiment and look at each person’s data, Voeks finding that her analysis was right. Participants transitioned from not giving CRs to giving them every time on a single trial. This would fit the cognitive picture of learning set out later by Brewer (1974) that subjects show insight into the procedure, consciously figuring out that CS predicts UR, and then producing the appropriate behavior (Leahey, 1979).

激进行为主义者认同沃克斯的总体结论。将个体行为归结为群体行为，可能会忽视心理学，无意中将其与其他社会科学融合。他们通常对一位或最多几位参与者进行实验，并发表个体行为的累积记录。斯金纳在其博士论文《有机体行为》（1938 年）中建立了这一模型。尽管标题暗示着对多种物种行为的调查，但该论文报告的是他用一只白鼠进行的广泛实验。在当今大数据时代，沃克斯和斯金纳的观点值得铭记。应对心理学（以及其他科学）重复性危机的一种策略是通过大幅增加研究参与者的数量来增强统计分析的有效性，但这只会让我们的注意力进一步偏离心理学的独特性：它不是一门社会科学，而是一门个人科学（Leahey，2007）。

Radical behaviorists agree with Voeks’ general conclusion. Aggregating individual behavior into group behavior risks losing sight of psychology, unwittingly melding it with the other social sciences. They typically perform their experiments on one or at most a few participants and publish individual cumulative records of individual behavior. Skinner established the model in his dissertation, Behavior of Organisms (1938). Despite the title suggesting a survey of behavior in many species, it reports his extensive experiments with one white rat. Voeks’ and Skinner’s point is important to remember in today’s era of Big Data. One strategy for dealing with the replication crisis in psychology (and other sciences) is to make statistical analysis more powerful by drastically increasing the number of participants studied, but this just draws our attention further away from psychology’s distinctiveness: It’s not a social science, but a personal science (Leahey, 2007.)

解读人类行为

Interpreting Human Behavior

像大多数心理学家一样，斯金纳希望利用心理学来改善人类生活，激进行为主义有一个专业组织，即^国际行为分析协会。7尽管斯金纳是一位实验心理学家，但他仍然致力于影响人类生活，尤其是在第二次世界大战之后。8他建造、宣传和分享了一种高科技婴儿床的计划，旨在减少母亲花在照顾婴儿上的时间（Skinner，1945 年）。他写了一本小说《瓦尔登湖 II》 ，描述了一个由受过科学训练的行为^控制者管理的乌托邦（Skinner，1948 年；参见 Rutherford，2009 年）。他在《超越自由和尊严》（Skinner，1971 年）和《关于行为主义》 （Skinner，1974 年）等书中普及了行为主义，并发表了关于现代社会缺点的演讲（例如，Skinner，1986 年）。

Like most psychologists, Skinner wanted to use psychology to improve human lives, and radical behaviorism has a professional organization, the Association for Behavior Analysis International.⁷ Although Skinner was an experimental psychologist, he nevertheless worked to influence human life, especially after World War II.⁸ He built, publicized, and shared plans for a sort of high-tech crib meant to reduce the amount of time mothers had to devote to infant care (Skinner, 1945). He wrote a novel, Walden II, describing a utopia run by scientifically trained behavior controllers (Skinner, 1948; see Rutherford, 2009). He popularized behaviorism in books such as Beyond Freedom and Dignity (Skinner, 1971) and About Behaviorism (Skinner, 1974), and he gave talks about the shortcomings of modern society (e.g., Skinner, 1986).

然而，斯金纳触及了其他科学心理学家回避的一个问题。哲学家威尔弗雷德·塞拉斯（Wilfred Sellars，1962）对人类的显性形象（即我们在日常生活中自我呈现的样子以及我们如何谈论自己）与人类的科学形象（即科学如何在科学研究的视角下构想人类）进行了颇具影响力的区分。现在应该清楚的是，这两种形象之间存在着巨大的张力，许多哲学家对此都进行了论述。一些人期待科学战胜民间心理学（例如，Dennett，1991，1996；但参见Leahey，1994；Stich，1983），他们认为民间心理学极其不准确，注定要失败。而其他人，包括塞拉斯本人，则认为这两种观点可以调和或融合（例如，Stich，1996）。少数人，尤其是塞尔（1994、1995、1997），认为科学的简化或消除是矫揉造作，且威胁甚微。不出所料，斯金纳属于第一阵营，他只承认日常常识性语言过于便捷，行为主义的语言无法将其消除。

Skinner, however, tackled one issue that other scientific psychologists avoid. Philosopher Wilfred Sellars (1962) has drawn an influential distinction between the manifest image of human beings—how we seem to ourselves and how we talk about ourselves in everyday life—and the scientific image of human beings—how science conceives of human beings under the lens of scientific research. It should be clear by now that there is enormous tension between these images, and many philosophers have written about it. Some look to a triumph of science over folk psychology (e.g., Dennett, 1991, 1996; but see Leahey, 1994; Stich, 1983), which they see as wildly inaccurate and hopelessly doomed. Others, including Sellars himself, believe the views can be reconciled or added together (e.g., Stich, 1996). A few, especially Searle (1994, 1995, 1997) view scientific reduction or elimination as pretentious and of little threat. Unsurprisingly, Skinner is in the first camp, conceding only that everyday common-sense language is too convenient to be eliminated by behaviorist speak.

斯金纳提出了一个令人震惊的论断：精神生活根本不存在。这引发了一个问题：为什么人们认为自己拥有精神生活，而实际上却没有？答案在于语言。他写道：“所有行为，无论是人类还是非人类，都是无意识的；当言语环境提供了自我观察所必需的偶然事件时，它就变成了‘有意识的’。”（斯金纳，1987，第782页）

Skinner makes the astonishing claim that there is no such thing as mental life, provoking the question why people think they have mental lives when they don’t. The answer lies in language. He writes, “All behavior, human or nonhuman is unconscious; it becomes ‘conscious’ when verbal environments provide the contingencies necessary for self-observation” (Skinner, 1987, p. 782).

斯金纳将《言语行为》（1957）视为他的杰作，书中他阐述了旨在论证上述主张的语言分析，特别是他的“语言技巧” （linguistic tact）概念。技巧旨在解决的首要问题是普遍指称，例如“猫”，而不是具体指称，例如“弗雷亚”（一只具体的猫）。指称行为主义通过刺激控制可以轻松处理细节问题。“弗雷亚”是一个由猫弗雷亚的正向辨别刺激控制的刺激。

Skinner regarded Verbal Behavior (1957) as his masterpiece, where he laid out the analysis of language meant to justify the claim above, specifically in his concept of the linguistic tact. The first problem the tact was meant to solve was the referent of universals, such as “cat,” as opposed to particulars, such as “Freya,” a particular cat. The referent of particulars could be handled easily in behaviorism with stimulus control. “Freya” is a stimulus controlled by the positive discriminative stimulus of the cat Freya.

普遍性更令人费解，因为不存在单一的物理CS+。根据源自亚里士多德的表征理论，普遍性是一种心理建构，普遍性术语可应用于此。然而，斯金纳运用了刺激控制理论，并根据个体的言语社群或文化对言语表达的塑造进行了修正。因此，孩子在猫面前说“猫”会得到差异化的强化，而在非猫面前说“猫”则不会得到强化，而是会因说“狗”或“牛”等其他操作词而得到强化。最终，“猫”会受到一类CS+的控制；我们不必说孩子已经获得了内在的“心理”概念。

Universals were more puzzling, because there is no single physical CS+. On the representational account descending from Aristotle, a universal is a mental construct to which universal terms apply. Nevertheless, Skinner applied stimulus control theory, modified by shaping of verbal utterances by a person’s verbal community, or culture. Thus, a child would be differentially reinforced for saying “cat” in the presence of cats, and not reinforced for saying “cat” in the presence of non-cats, but instead reinforced for other operants, such as “dog” or “cow.” Eventually “cat” comes under the control of a class of CS+s; we do not need to say the child has acquired an inner “mental” concept.

斯金纳运用“触觉”来解释我们对精神生活的信念。我们可以触觉——即谈论，或者用行为语言来表达——我们自己的行为和私人刺激。在触觉私人刺激的情况下，感觉它们在我们体内，加上他人无法感知它们，这创造了一种内省私人精神世界的假象，而实际上，我们只是在接触我们身体的私人内部环境。在触觉自身行为的情况下，我们实际上与外部观察者处于相同的位置，但我们被自己的民间心理学误导了。想象一下，在一个寒冷的除夕夜，你参加了一个聚会，你在主人的衣橱里戴帽子、外套和手套。主人走过来问你在做什么，你回答说：“很晚了，我要走了。”现在，主人可以从之前的事件和你的行为推断出这一点。斯金纳认为，事实上，你也在做同样的事情。你的行为是由先前的刺激（球落地）和当前刺激（你在壁橱旁）引起的，现在，在这种刺激的提示下，你触碰了自己穿上冬装的行为，并像你的主人一样发现你现在要离开了，因为过去做过这些事之后，必然会出现离开的行为。按照斯金纳的说法，只有当你触碰了这些行为的原因时，你才会意识到这些原因，但触碰本身并没有引发你的行为，它只是描述了行为的先例。

Skinner deployed the tact to explain our belief in mental life. We can tact—i.e., talk about, or in behavior-speak make verbal responses to—our own behavior and private stimuli. In the case of tacting private stimuli, the feeling that they are inside us plus the fact that others don’t sense them creates the illusion of introspection of a private mental world when, in fact, we are merely accessing the private inner environment of our body. In the case of tacting our behavior, we are really in the same position as outside observers, but we have been misled by our folk psychology. Imagine you’ve been at a party on a cold New Year’s Eve, and you are at your host’s closet putting on your hat, coat, and gloves. Your host comes by and asks what you are doing, and you reply, “It’s late and I’m leaving now.” Now your host can infer this from previous events and your behavior. Skinner argues that you, in fact, do the same thing. Your behavior has been caused by previous (the ball has dropped) and current (you are at the closet) stimuli, and now, prompted to do so, you tact your own behavior of putting on your winter gear and discover, as your host has, that you are leaving now, because doing these things in the past has been reliably followed by leaving behavior. You become conscious, in Skinner’s sense, of the causes of your behavior only when you tact them, but the tacting did not cause the behavior, it simply described its antecedents.

拥有精神生活是一种错觉，它源于我们能够用言语表达公共和私人刺激以及我们自身观察到的行为，而古代哲学家传承下来的错误民间心理学则助长了这种错觉。“有谁解释过心灵如何作用于身体，或身体如何作用于心灵？这类问题根本不应该问。”（Skinner，1987，第785页）

Having a mental life is a delusion brought about by our ability to verbalize public and private stimuli and our own observed behavior, abetted by an erroneous folk psychology descended to us from ancient philosophers. “[H]as anyone ever explained how the mind works on the body or the body on the mind? Questions of that sort should never have been asked” (Skinner, 1987, p. 785).

巴甫洛夫学习的现代联想理论

Modern Associative Theory of Pavlovian Learning

尽管已经提出了一些超强符号系统条件反射理论（例如，Holyoak、Koh 和 Nisbett，1989；Boddez 等人，2014），但大多数现代理论都是联想性的，经常与神经科学（Honey、Iordanova 和 Good，2014；Houwer，2011）或神经网络（Mondragón、Alonso 和 Kokkola，2017）相融合。

Although some overpowered symbol system theories of conditioning have been proposed (e.g., Holyoak, Koh, & Nisbett, 1989; Boddez et al., 2014), the majority of modern theories are associative, often blending with neuroscience (Honey, Iordanova, & Good, 2014; Houwer, 2011) or neural nets (Mondragón, Alonso, & Kokkola, 2017).

我们现在的情况

Where We Are Now

我们追踪了行为理论的发展，以及它为认知科学提出的至今仍活跃的理论问题，尽管早期的理论家主要研究的是动物行为。现在我们来谈谈信息处理认知科学的研究。尽管信息处理理论是新兴的——它用计算机来比喻人类——但我们会发现，随着新问题的出现，相同的理论问题依然存在。

We have tracked the development of behavioral theory and the still active theoretical issues it raises for cognitive science, even though the older theorists primarily investigated animal behavior. Now we turn to the study of information processing cognitive science. Although information processing theory is new—its metaphor for the human is the computer—we will find the same theoretical problems remain as new ones are added.

外卖

Takeaway

虽然激进行为主义作为认知科学理论并不令人满意，但应用行为分析为改变行为提供了有效的工具包。我们已经了解了两个重要的观点，即有机体总是对的，以及你必须从有机体的行为开始。第三个是强化相对论，非正式地称为普雷马克原理（Premack，1965）。它是强化思维重要转变的一部分。在 SR 学习理论中，强化物被认为是与身体需求相关的东西，例如食物——饥饿或水——口渴。然而，在新的概念中，强化物是活动：吃是强化物，而不是食物；喝是强化物，而不是水。此外，强化行为不一定与生物驱动力有关。这种变化具有重要的理论和实践意义（Baum，2012）。

While radical behaviorism is unsatisfactory as a theory in cognitive science, applied behavior analysis provides an effective toolkit for changing behavior. We’ve already met two important ideas, that the organism is always right and that you must begin with the behaviors the organism has. A third is the relativity of reinforcers, known informally as the Premack Principle (Premack, 1965). It is part of an important shift in thinking about reinforcement. In S-R theories of learning, reinforcers were thought of as things tied to physical needs, such as food – hunger or water – thirst. However, in the new conception, reinforcers are activities: eating is the reinforcer, not food; drinking is the reinforcer, not water. Moreover, reinforcing behaviors are not necessarily tied to biological drives. This change has important theoretical and practical implications (Baum, 2012).

如果我们仅仅观察动物或人在其自然环境中的自发行为，我们就会发现某些活动比其他活动进行得更频繁。如果我们将这些活动绘制成图表，我们就可以为任何生物体建立一个从最多做到最少做的活动的层次结构。经济学家在 RCT 中将这些称为偏好层次结构，但斯金纳主义者不喜欢心理主义术语。普雷马克原理认为，对于任何两项活动，我们可以通过让更受欢迎行为的执行取决于其表现来增加较不受欢迎行为的频率。例如，在斯金纳箱实验开始时，我们让进食取决于按压杆。或者，我们可以让进食取决于向左转圈。

If we simply observe an animal or person spontaneously behaving in its natural environment, we will see that some activities are performed more frequently than others. If we chart these, for any individual organism we can establish a hierarchy of activities from most done to least done. Economists call these preference hierarchies in RCT, but Skinnerians don’t like mentalistic terms. The Premack Principle says that for any two activities, we can increase the frequency of a less preferred behavior by making access to a more preferred behavior contingent on its performance. For example, at the beginning of an experiment in a Skinner box, we make eating contingent on bar-pressing. Or we could make eating contingent on turning circles to the left.

这一修正观点的第一个重要含义是扩展了强化刺激的种类。老鼠在活动轮中会自发奔跑，因此我们可以将进入活动轮的条件设定为按下杆子打开轮子的门。第二个含义是，个体的偏好层次存在差异，因此没有放之四海而皆准的强化模式。例如，老师会发现有些学生喜欢安静地阅读而不是参加体育运动，而有些学生则相反。我们可以将阅读时间设定为玩耍时间，从而提高阅读者的积极性，而对于运动爱好者则相反（例如，参见 Geiger，1996）。

The first important implication of this revised view is expanding the kit of reinforcers. Rats run spontaneously in their activity wheels, so we could make access to the wheel contingent on bar-pressing to open a door to the wheel. The second implication is that there are individual differences in preference hierarchies, so there is no one-size-fits-all template of reinforcement. For example, teachers will find students who prefer reading quietly to participating in sports, and others for whom the opposite is true. One can get the reader more active by making reading time contingent on play, and the reverse for the sports enthusiast (see e.g., Geiger, 1996).

我在物理康复期间对自己进行了行为分析。正如我之前提到的，站立比行走更难（对我来说，也更痛苦）。然而，我的治疗师希望按照逻辑顺序教授技能，先站立，再行走。所以，我和治疗师达成了一项协议（并试图传达行为方面的见解）：如果我尝试（塑形）站立，他们就会把我拉起来，让我走几步（强化），随着站立能力的提高，我会走更多路。现在我又能站起来了，我仍然把散步作为强化物。

I applied behavior analysis to myself during physical rehab. As I mentioned before, standing up is harder (and for me then, much more painful) than walking. However, my therapists wanted to teach skills in the logical sequence, standing before walking. So, I made a deal (and tried to pass along behavioral insights): If I tried a bit (shaping) to stand, they’d then haul me up and let me walk some steps (reinforcement), with more walking as standing improved. Now that I’m on my feet again, I still use taking a walk as a reinforcer.

阅读材料

Readings

关于应用行为方法的书籍有很多，但最有趣的是艾米·萨瑟兰的《沙姆教给我的人生、爱情和婚姻：动物及其训练师给人类的启示》（纽约：兰登书屋，2006年）。最初，她只是在《纽约时报》上发表一篇关于海洋世界动物训练的文章，并将它应用到丈夫身上，这让她一夜成名，并签下了书籍合同和电影合约（据我所知，这份合约从未兑现）。

There are lots of books on applied behavior methods, but the most fun is Amy Sutherland’s What Shamu Taught Me about Life, Love, and Marriage: Lessons for People from Animals and Their Trainers. (New York: Random House, 2006). What started out as a New York Times piece on Sea World animal training that she applied to her husband led to media stardom, a book deal, and a movie contract (never fulfilled as far as I can tell).

电影

Movies

《心理学史的进展》中有一个关于巴甫洛夫条件反射的俄罗斯汇编电影的链接：https://ahp.apps01.yorku.ca/2010/02/pavlovian-conditioning-on-film-2/。

Advances in the History of Psychology has a link to a Russian-assembled film on Pavlovian conditioning: https://ahp.apps01.yorku.ca/2010/02/pavlovian-conditioning-on-film-2/.

行为分析师推荐了一些电影：https://www.appliedbehavioranalysisprograms.com/lists/5-great-movies-for-applied-behavior-analysts/。

Behavior analysts have a short list of recommended movies: https://www.appliedbehavioranalysisprograms.com/lists/5-great-movies-for-applied-behavior-analysts/.

《办公室》第三季第15集（https://www.nbc.com/the-office/video/jims-pavlovian-prank-on-dwight-the-office/4141507 ）中有一幕，将自己描述为巴甫洛夫条件反射练习，但其呈现方式具有误导性。尝试识别US、CS、UR和CR，但要小心！同时，还有未言明的操作性学习在进行。操作性学习总是存在的：经过几天的条件反射训练后，当巴甫洛夫的狗早上从睡笼中被释放出来时，它们会急切地跑到走廊尽头的实验室，跳上设备，准备接受强化。识别《办公室》场景中操作性学习的元素。

A sequence in season 3, episode 15 of The Office (https://www.nbc.com/the-office/video/jims-pavlovian-prank-on-dwight-the-office/4141507) presents itself as an exercise in Pavlovian conditioning, but its presentation is misleading. Try to identify the US, CS, UR, and CR, but be careful! There’s also—unspoken—operant learning going on at the same time. There always is: After a few days of conditioning, when Pavlov’s dogs were released from their sleeping cages in the morning, they eagerly ran down the hall to the lab and jumped onto the equipment, ready to be reinforced. Identify the elements of operant learning in The Office scenario.

人类信息处理系统的模态模型

The Modal Model of the Human Information Processing System

第二章区分了串行和并行信息处理。长期以来，正如奈瑟尔的观点，串行处理主导了认知心理学和人工智能，直到联结主义的复兴。串行处理认知模型占据主导地位的一个重要原因是，意识是一个串行处理器，每次计算一步，我们很容易像亚里士多德那样，将感知视为一个串行过程，从受体对刺激的感知、早期感知（特殊感官）到意识中对物体的识别（常识），再到直接的想象存储和记忆。图5.1展示了最常被引用的“模态模型”的示意图。（Baddeley、Hitch & Allen，2019）信息处理理论（Atkinson and Shiffrin，1968、1971）在许多方面与亚里士多德的理论相似。

Chapter 2 distinguished between serial and parallel information processing. For a long time, as in Neisser, serial processing dominated both cognitive psychology and artificial intelligence until the revival of connectionism. An important reason for the dominance of serial processing models of cognition is that consciousness is a serial processor, computing one step at a time, and it’s easy to think of perception as a serial process, as Aristotle did, moving from sensation of stimuli by receptors, early perception (the special senses), and recognition of objects in consciousness (common sense) to immediate imaginal storage to memory. Figure 5.1 shows a diagram from the most cited, “modal model” (Baddeley, Hitch, & Allen, 2019), of information processing (Atkinson and Shiffrin, 1968, 1971), which in many ways resembles Aristotle’s.

虽然这个模型如今已过时，但它仍然提供了一个有用的起点。它直观易懂，而且写作和教学必然是连续进行的。因此，我们将以此图的结构来指导我们的讨论，尽管最终我们会发现，并行处理是人类思维的常态，而非例外。

Although this model is now dated, it nevertheless provides a useful place to start. It’s intuitive, and, moreover, writing and teaching are necessarily serial enterprises. Thus, we will use this figure’s structure to guide our discussion, although by the end we will see that parallel processing is the rule in the human mind, not the exception.

在我们继续之前，您应该观看以下视频，其中演示了一种称为“注意力盲视”的现象：

Before we proceed, you should watch the following videos demonstrating a phenomenon called inattentional blindness:

1. https://www.youtube.com/watch?v=vJG698U2Mvo . ¹
2. https://www.youtube.com/watch?v=vJG698U2Mvo.¹
3. https://www.youtube.com/watch?v=IGQmdoK_ZfY。
4. https://www.youtube.com/watch?v=IGQmdoK_ZfY.

认知水平框架

Cognitive Level Frameworks

让我们首先来看看 Marr 提出的两个用于研究早期信息处理的认知层面框架。这些框架并非关于心智如何从刺激到意识处理信息的实际理论，而是用于构建理论的框架：心智需要解决哪些适应性问题？哪些方法可以算作解决这些问题的充分和/或最佳方法？人类或人工智能心智需要哪些认知资源？这些框架还具有其他优势，可以应用于认知科学的其他领域，例如思维和决策。

Let’s begin by looking at two Marr cognitive-level frameworks for studying early information processing. These are not theories about how the mind actually processes information from stimulus to consciousness, but frameworks for building theories: What adaptive problem does the mind have to solve? What might count as adequate and/or optimal ways of solving these problems? What cognitive resources does the human or AI mind need? These frameworks have the additional virtue of applying in other areas of cognitive science, such as thinking and decision making.

镜头模型

The Lens Model

感知的透镜模型由匈牙利心理学家埃贡·布伦斯维克（Egon Brunswik，1903-1955）提出，他曾与托尔曼（Tolman）合作（Leary，1987）。他的模型概括了感知的表征理论。布伦斯维克（Brunswick，1952）提出，感知涉及根据环境中物体（感知的远端物体）提供的近端线索构建内部表征（感知的近端物体）。感知的成功取决于近端物体与远端物体的匹配程度（图5.2）。

The lens model of perception was developed by Egon Brunswik (1903–1955), a Hungarian psychologist who worked with Tolman (Leary, 1987). His model adumbrates the representational theory of perception. Brunswick (1952) proposed that perception involved constructing an internal representation (the proximate object of perception) from proximal cues supplied by the object in the environment (the distal object of perception). Perception succeeds to the extent that the proximate object matches the distal object (Figure 5.2).

在透镜模型中，感知是信息整合这一更广泛的认知过程的一个实例。就像凶杀案调查中的侦探一样，我们掌握着一组离散的事实，必须判断哪些相关、哪些不相关、哪些最重要，并解析模棱两可的线索，等等，直到它们指向凶手。

In the lens model, perception is an instance of a broader set of cognitive processes called information integration. Like a detective in a murder investigation, we find ourselves in possession of a discrete set of facts and must decide which are relevant and which are not, which are most important, resolve ambiguous clues, and so on until they point to the killer.

信号检测模型

The Signal Detection Model

我们的另一个框架，信号检测理论(SDT) (Green & Swets, 1966; Wickens, 2002)，则持不同观点。它既不涉及表征哲学，也不涉及实在论哲学，而是将感知视为基于经验的判断，而非构建世界图景的过程。就我们的目的而言，信号检测理论最重要的一点在于它将行动融入了感知图景。感知并不会随着近端对象的创建而结束，而是在我们根据已得出的结论采取行动时继续进行。根据线索，我们得出结论，是罗科杀死了弗内博恩。说出他的名字会有什么后果？如果他有罪呢？如果他无辜呢？我们或许确信他有罪，但我们应该审判他吗？

Our other framework, signal detection theory (SDT) (Green & Swets, 1966; Wickens, 2002), takes a different view. It doesn’t implicate either representational or realist philosophies of perception, seeing perception as a judgment passed on experience rather than a process of building up a picture of the world. For our purposes what’s most important about signal detection theory is that it incorporates action into its picture of perception. Perception does not end with the creation of the proximal object but continues as we act on the conclusion we have reached. From clues we conclude that Rocco killed Fonebone. What are the consequences of naming him? What if he is guilty? What if he is innocent? We may be convinced he is guilty, but should we put him on trial?

声纳探测（SDT）的典型案例是声纳操作员，就像电影《猎杀红色十月》中描述的那样。声纳操作员通过拖曳式麦克风阵列聆听海面的声音，听到各种各样的声音，但其中只有一种可能很重要：敌方潜艇的声音。操作员的工作是在海洋背景噪音中判断是否存在潜艇信号（图5.3）。

The paradigm case of SDT is a sonar operator, as depicted in movies such as Hunt for Red October. The sonar operator listens to the sounds of the sea transmitted to him by a towed array of microphones, hearing a myriad of noises, only one of which might be important: The sound of an enemy submarine. The operator’s job is to determine if there is the presence of a submarine signal against the background noise of the ocean (Figure 5.3).

与感知的哲学理论不同，SDT 是由工程师开发的，它考虑了得出有信号或没有信号的结论的后果。有四种可能的结果：有信号时检测到信号；没有信号时“检测到”信号；有信号时未检测到信号；没有信号时未“检测到”信号。图 5.4总结了这些情况，并为我们假设的潜艇狩猎场景做了注释。图中单元格中的垂直线表示条件概率，例如 P(Y|N) 表示当只有噪声存在时说“是”的概率，而 P(Y|SN) 表示在信号和噪声同时存在的情况下说“是”的概率。完美的声纳操作员将具有以下接收器操作特性(ROC)：P(Y|N) = 0（误报）；P(Y|S) = 1（命中）； P(N|N) = 1（正确拒绝），P(N|SN) = 0（未通过）。

Unlike philosophical theories of perception, SDT was developed by engineers, and it considers the consequences of concluding there is a signal or there isn’t a signal. There are four possible outcomes: Detecting a signal when there is one; “detecting” a signal when there isn’t one; failing to detect a signal when there is one; and not “detecting” a signal when there isn’t one. These cases are summarized in Figure 5.4, glossed for our hypothetical submarine-hunting scenario. The vertical lines in the cells of the figure indicate conditional probabilities, e.g., P(Y|N) means the probability of saying yes when only noise is present, and P(Y|SN) means the probability of saying yes given that signal and noise are both present. A perfect sonar operator would have the following receiver-operating characteristics (ROC): P(Y|N) = 0 (false alarm); P(Y|S) = 1 (hit); P(N|N) = 1 (correct rejection), and P(N|SN) = 0 (miss).

在SDT的框架下，有两个因素决定了感知过程的结果。第一个因素是d'，即敏感度，它与确定刺激是否存在的难度有关；第二个因素是标准设定，它与感知者给出正确答案的重要性有关。

Given the framework of SDT, two considerations determine the outcome of perceptual processes. The first, d’ or sensitivity, concerns the difficulty of determining whether or not a stimulus is present, and the second, criterion setting, concerns how important it is that the perceiver give the right answer.

在某些情况下，信号在背景噪声中并不突出（图 5.5）。

In some cases, a signal does not stand out well against background noise (Figure 5.5).

想象一下，你有个朋友跑马拉松，他请你为他加油，让他参加一场金刚主题的马拉松。你应该为谁加油，或者给谁倒水？在一群穿着大猩猩服的跑步者（噪音）中找出你的朋友（信号）可不是件容易的事。

Imagine that you have a friend who runs marathons, and he asks you to root for him in a King Kong–themed marathon. Who should you root for or give water to? Picking out your friend (signal) among all the gorilla-suited runners (noise) is difficult.

另一天，另一场比赛可能会更容易（图 5.6）。

Another day, a different race might be easier (Figure 5.6).

其次，你所看到的，或者你认为自己看到的，不仅取决于环境带来的挑战，还取决于你如何根据情境的重要性进行感知。例如，在《隐形大猩猩》中实验中，你的任务是计算传球次数，因此你设定自己将这些次数视为对抗视频视觉噪声的信号。传球检测比大猩猩检测更重要（图 5.7）。

Second, what you see, or think you see, is determined not just by the challenges posed by the environment, but also by how you have disposed yourself to perceive in accord with the importance of the situation. For example, in the invisible gorilla experiment, your job is to count the number of ball passes, and so you set yourself to see them as the signal against the visual noise of the video. Ball pass detection is more important than gorilla detection (Figure 5.7).

如果你接受了计数传球的任务，那么传球就是需要检测的信号，而场景的其余部分则是噪音，所以你很可能看不到也检测不到大猩猩，即使它的图像出现在你的视网膜上。另一方面，如果你被要求计数动物传球给球员，那么大猩猩就是信号，球员就是噪音，你会看到大猩猩，但不知道传球了多少次。

If you have accepted the assigned task of pass-counting, the passes are the signals to detect and the rest of the scene is noise, and so you may very well fail to see and detect the gorilla, even though its image crosses your retina. On the other hand, should you be told to count animals passing the players, the gorilla would be the signal and the players the noise, and you would see the gorilla but have no clue how many ball passes were made.

重要的是要理解，SDT 并不声称刚才描述的过程是有意识的。SDT 认为感知包含推理，但这些推理是无意识的，正如赫尔曼·冯·亥姆霍兹（1821-1894）在 19 世纪首次提出的那样（Helmholtz, 1867; Hatfield, 2002）。因为这并不重要，所以你实际上并没有看到大猩猩。

It is important to understand that SDT does not claim that the process just described is conscious. SDT says that perception involves inferences, but that these inferences are unconscious, as proposed first by Herman von Helmholtz (1821–1894) in the 19th century (Helmholtz, 1867; Hatfield, 2002). Because it’s not important, you genuinely do not see the gorilla.

Drew、Vo 和 Wolf (2013) 强调了这一结论。他们研究了放射科医生在肺部 CAT 扫描中识别跳舞大猩猩照片的能力。他们要求放射科医生和非专业参与者检查一系列 CAT 扫描图像，寻找并计数被称为结节的异常。其中一张扫描图像包含一张黑白图像，其中一只跳舞的大猩猩大约有火柴盒大小；它清晰可见；实验结束时，当参与者被特别询问目标图像是否有任何异常时，每个人都指出了大猩猩。然而，在寻找结节的过程中，没有一位天真的参与者报告了大猩猩，24 位放射科医生中只有 4 位报告了大猩猩。Drew、Vo 和 Wolf 还使用眼动追踪设备记录了每位参与者视网膜扫描的区域，发现即使直视大猩猩，也很少报告。这进一步证实了他们的结论：大猩猩并非被体验到，而是被看到但被忽略了。在《血字的研究》中，夏洛克·福尔摩斯责备华生：“你只看不观察。”

This conclusion is underscored by Drew, Vo, and Wolf (2013), who looked at radiologists’ ability to see photos of a dancing gorilla in CAT scans of lungs. They asked radiologists and lay participants to examine a series of CAT scans looking for and counting abnormalities called nodules. One of the scans included a black-and-white image of a dancing gorilla about the size of a matchbox; it was easily visible; when at the end of the experiment, participants were specifically asked if there was anything odd about the target image, everyone pointed out the gorilla. However, while hunting for nodules, none of the naive participants and only four of 24 radiologists reported the gorilla. Drew, Vo, and Wolf also used an eye-tracking device to record the areas each participant’s retina scanned, finding that the gorilla was rarely reported even when directly looked at, reinforcing the conclusion that it was not experienced, rather than seen but ignored. In A Study in Scarlet, Sherlock Holmes upbraids Watson: “You see but you do not observe.”

信息处理的类型

Types of Information Processing

信息处理有多种类型。本章重点在于理解两对不同的处理方式。这两种认知过程不应被视为非此即彼，即一个过程在没有另一个过程的情况下发生。两者都始终在进行。弄清楚它们的相对贡献是一个实证问题，正如第一个对比所展示的那样。

There are various types of information processing. For this chapter, understanding two pairs of distinct processing is important. Neither of these cognitive processes should be thought of as either/or, one occurring in the absence of the other. Both go on all the time. Figuring out their relative contributions is an empirical question, as the demonstration of the first contrast shows.

自上而下与自下而上的处理

Top-Down vs. Bottom-Up Processing

自上而下处理和自下而上处理之间的区别很重要，但很容易被误解，尤其是在将思维视为一个串行处理器，先处理一个过程，然后再处理另一个过程的情况下。此外，认知科学中沿用至今的标签不如最初提出的其他标签那样具有描述性，这些标签最初是为了捕捉这种区别而提出的。我更喜欢用更具描述性的术语“数据驱动处理”来描述自下而上的处理，用“概念驱动处理”来描述自上而下的处理。总而言之，其理念是：经验（以及我们的记忆）是经验形成过程中获得的实际事实（数据驱动/自下而上）与我们在经验形成过程中对世界的期望、信念、价值观和积累的认知（概念驱动/自上而下）的混合体。隐形大猩猩实验已经让我们初步领略了这一理念，它是本章其余部分的基础。

The distinction between top-down processing and bottom-up processing is important but easy to misunderstand, especially if one thinks of the mind as a serial processor that first does one and then the other. Moreover, the labels that have stuck in cognitive science are not as descriptive as other ones originally proposed to capture the distinction. I prefer the more descriptive terms data-driven processing for bottom-up processing and conceptually-driven processing for top-down processing. Put together, the idea is that experience (and our memories) are mixtures of the actual facts (data-driven/bottom-up) obtaining in the world during the formation of an experience and the expectations, beliefs, values, and accumulated acquaintance (conceptually driven/top-down) with the world that we bring to the formation of an experience. The invisible gorilla experiments already give us a taste of this idea, and it’s fundamental to the rest of the chapter.

这两种类型的处理方式以一种非此即彼的方式构成了早期感知哲学的基础。复制理论和实在论认为，经验强加于心灵，因此，我们看到的是世界事实的本来面目。唯心论认为感知存在不可挽回的缺陷，坚持认为我们已经拥有了真理，这些真理是与生俱来的或可以通过理性证明的概念。亚里士多德的经验主义与实在论相符，直到特殊感官记录的事实被常识所判断。然而，亚里士多德的图景假设了我们现在知道是错误的东西，即在概念应用于数据时存在一个判断的时刻，在那个时刻，我们对世界的感官图景，无论准确与否，都会结晶成一个有意义的感知。正如我们将看到的，自下而上和数据驱动的过程从感知的最初几毫秒开始，一直持续到数年乃至数十年的记忆存储和检索。对于每一种经验和每一段记忆，研究人员的诀窍在于解开它们（图5.8）。

In an either/or way, these two types of processing underlie earlier philosophies of perception. Copy theory and realism say that experience imposes itself on the mind and, therefore, we see the facts of the world as they are. Idealism dismisses perception as irretrievably flawed, insisting that we already possess the truth as innately given or rationally provable concepts. Aristotle’s empiricism agrees with realism until the facts recorded by the special senses are judged by common sense. However, Aristotle’s picture assumes something we now know is wrong, that there is a single moment of judgment when concepts are applied to data, a single moment when our sensory picture of the world, accurate or not, crystallizes into a meaningful perception. As we will see, bottom-up and data-driven processes operate from the first milliseconds of sensation through years and decades of memory storage and retrieval. For every experience and every memory, the trick for researchers is untangling them (Figure 5.8).

探究感觉记忆

Investigating Sensory Memory

冯特受到直觉模型的联想版本的影响。在联想主义中，诸如CAT之类的有意义的复杂概念是由更简单、意义更不明晰的元素构成的，类似于视频像素（参见模式识别部分）。在《观念之路》中，提出了一个问题：在任何给定的时间点，意识可以容纳多少个元素。在冯特之前，哲学家们就此争论了数十年，但他们未能得出结论，因为他们的空想内省并非实验性的。控制。冯特构建了一个标准化的情况，参与者可以对他们所看到的内容做出简单、直接的回答。

Wundt was influenced by an associative version of the intuitive model. In associationism, meaningful complex ideas, such as CAT, were made up of simpler, less meaningful elements similar to video pixels (see section on pattern recognition). In the Way of Ideas the question arose as to how many elements consciousness could hold in any given moment in time. Philosophers argued about this for decades before Wundt, but they had not reached a conclusion because their armchair introspections were not experimentally controlled. Wundt constructed a standardized situation to which participants could respond with simple, immediate answers about what they saw.

冯特使用一种名为速视仪的装置，在极短的时间内（不到一秒）向受试者呈现简单的视觉刺激。如今，计算机显示器已用于此目的，但早期的研究人员使用了各种机械和机电设备。以下是冯特刺激阵列的示例（图 5.9）。

Wundt used a device called a tachistoscope to present simple visual stimuli to participants for very brief—less than a second—exposures. Today, computer displays are used for this purpose, but earlier researchers used a variety of mechanical and electro-mechanical devices. Here’s an example of one of Wundt’s stimulus arrays (Figure 5.9).

在最初的空白屏幕中央有一个小点，这是参与者关注的注视点。然后字母闪烁一小会儿，参与者报告尽可能多的字母。因为他要求参与者从整个 16 个字母阵列中报告尽可能多的字母，所以冯特的方法被称为整体报告程序。冯特发现，天真的参与者可以报告大约三到四个字母，这是一个至今仍然可靠的发现。在冯特的时代，参与者不是心理学入门学生（当时没有），而是成为熟练的内省者的教授和研究生。在这个实验中，熟练的参与者可以报告多达六个字母。最后，所有参与者都认识到字母数超过三到六个，但是他们看得不清楚和不清楚，所以无法报告。

At the center of an initial blank screen was a small dot, the fixation point the participant focused on. Then the letter display was flashed for a fraction of a second, and the participant reported as many letters as possible. Because he asked for participants to report as many letters as possible from the whole array of 16, Wundt’s method is called the whole report procedure. Wundt found that naïve participants could report about three to four letters, a robust finding to today. In Wundt’s day, participants were not intro psych students (there being none), but fellow professors and grad students who became practiced introspectors. In this experiment, practiced participants could report up to six letters. Finally, all participants recognized that there were more than three to six letters, but these were not seen clearly and distinctly, and so they could not be reported.

从现代视角来看，冯特“发现”了感觉记忆，但他用注意力而非记忆来解释他的发现。冯特预见了后来注意力的“聚光灯”理论（Posner, Snyder, & Davidson, 1980），他将视觉意识视为一个领域，其中存在一个中心的、被关注的焦点区域，周围环绕着一个更大、甚至难以界定的边缘区域（图5.10）。

From a modern perspective, Wundt had “discovered” sensory memory, but he explained his findings in terms of attention rather than memory. Anticipating later “spotlight” theories of attention (Posner, Snyder, & Davidson, 1980), Wundt regarded visual consciousness as a field in which there was a central, attended focal area surrounded by a larger, indeed undefinably large, fringe area (Figure 5.10).

冯特的结论受到了斯珀林（1960）的质疑。斯珀林同意冯特的结果，修改了他的方法，并提出了一种非常早期、语义原始的记忆形式，称为“图像”。他认为，整个报告过程不能排除这样一种可能性，即即使字母阵列消失后，参与者仍可以暂时看到整个阵列，因为报告看到了多少个字母需要时间，在此期间图像衰减为虚无。为了支持他的主张，斯珀林发明了部分报告程序，其中提示参与者只回忆显示屏上的一行。斯珀林实验中的显示屏是 3 x 4 的字母显示屏，随后是随机选择的高、中或低音调。如果他们听到高音，参与者要报告最上面一行；如果是中音，则报告中间一行；如果是低音，则报告最下面一行。阵列偏移和提示音呈现之间的间隔从立即到大约一秒不等^。2

Wundt’s conclusions were challenged by Sperling (1960), who agreed with Wundt’s results, modified his method, and proposed the existence of a very early, semantically primitive, form of memory called the icon. He argued that the whole report procedure could not rule out the possibility that the entire letter array was temporarily available to participants even after it disappeared because it took time to report how many letters were seen, during which period the image decayed to nothingness. To support his claim, Sperling invented the partial report procedure in which the participant was cued to recall just one row of the display. The display in Sperling’s experiment was a 3 x 4 display of letters followed by a randomly chosen high, medium, or low-pitched tone. If they heard the high tone, participants were to report the top row; if the medium tone, the middle row; and if the low tone, the bottom row. The interval between the offset of the array and the presentation of the cuing tone was varied from immediate to about a second.²

斯珀林发现，如果在阵列之后立即或几乎立即呈现音调，参与者可以报告提示行，但不能报告其他行。随着阵列偏移和音调之间的时间变长，性能下降，直到大约四分之三秒时，部分报告和完整报告相同。斯珀林的结论是，在那短暂的时间内，参与者拥有一个图像，即字母阵列的图标，他们可以看到并用它来报告提示行。然而，正如他在批评冯特的整体报告程序中所暗示的那样，图标会自发衰减到没有。他还表明，在阵列偏移后立即呈现刺激物可以擦除或掩盖图标，甚至是图标的碎片。例如，如果在字母刚刚出现的位置呈现一个比字母稍大的环，参与者报告只看到了环，而不是被环包围的字母。如果在整个阵列所在的位置呈现视觉白噪声，参与者根本不会报告字母。掩蔽对于研究无意识感知的心理学家来说很有用，因为他们可以呈现刺激但掩盖它们，并观察它们是否仍然影响行为。

Sperling found that if the tone was presented immediately or nearly immediately after the array, participants could report the cued row, but not the other rows. As the time between array offset and tone got longer, performance declined until at about three-quarters of a second, partial reports and whole reports were the same. Sperling concluded that for that brief time participants possessed an image, an icon of the letter array, that they could look at and use to report the cued row. However, as suggested in his criticism of Wundt’s whole report procedure, the icon spontaneously decayed to nothing. He also showed that the icon, or even pieces of the icon, could be erased, or masked, by the presentation of stimuli just after the array offset. For example, if a ring slightly bigger than a letter was presented where a letter had just been, participants reported seeing the ring alone rather than the letter surrounded by the ring. If visual white noise was presented where the whole array had been, participants never reported letters at all. Masking would prove useful to psychologists interested in unconscious perception, because they could present stimuli but mask them, and see if they nevertheless affected behavior.

感觉记忆的特点

Characteristics of Sensory Memory

感觉记忆容量巨大，能够短暂地容纳远超意识所能掌握的信息。它容易发生两种形式的遗忘：衰减，以及新信息的干扰，例如掩蔽效应。

Sensory memory has a large capacity, being able to briefly hold much more information than can be grasped by consciousness. It is subject to two forms of forgetting, decay, and interference by the arrival of new information, as seen in masking.

关于感觉记忆，最具争议的在于它包含什么信息以及这些信息是如何构成的。当它在信息处理的串行处理模型中首次被提出时，它被认为是一种前分类存储。也就是说，这些字母还没有被归类为字母——它们只是白底黑字的图案。分类被认为发生在模式识别的处理流的后期。然而，这种假设已被证明是错误的。例如，Merikle（1980）展示了随机混合字母和数字的数组，并提示参与者在某些情况下按行报告，或者在其他情况下按字母或数字报告，结果发现两种情况下的表现都一样好。如果图标是一种真正的前分类存储，参与者就不应该能够按语义类别对他们的报告进行排序。

What became most controversial about sensory memory is what information it contains and how that information is structured. When it was first proposed in the context of the serial processing model of information processing, it was thought to be a pre-categorical store. That is, the letters had not yet become, i.e., categorized as, letters—they were just patterns of black on white. Categorization was thought to take place later in the stream of processing during pattern recognition. This assumption, however, has proved wrong. For example, Merikle (1980) presented arrays that randomly mixed letters and numbers, and cued participants to report by rows in some cases or by letter versus number in others, finding that performance was equally good in both cases. If the icon was a true pre-categorical store, participants should not have been able to sort their reports by semantic category.

除了感觉记忆并非斯珀林所想的那么简单之外，其确切性质仍存在争议（Jolicoeur、Lefebvre & Martinez-Trujillo，2016）。一个有趣的假设与预测处理相关，即图像由多层构成，每一层由不同的神经过程创建，并包含不同类型的信息（Rensink，2014；类似模型参见Capiello & Zhang，2016）。例如，最先创建的、最简单的层实际上是未经处理的物理刺激表征，它会迅速衰减。另一层是在经过一些概念处理后创建的，由神经回路从大脑高级区域反馈到视网膜的输入回路，正是这一层使得梅里克尔（Merikle）发现的感觉阵列能够进行概念处理。另一方面，Treisman & Lages（2013）认为，图像和感觉记忆根本不存在于信息处理的某个阶段。他们利用 SDT 提出，生物体会调整其决策标准以最佳地适应其感官世界，并且随着决策标准的调整，可以观察到感官输入的不同方面，无论是物理的还是语义的。

Beyond the fact that sensory memory is not as simple as Sperling thought, its exact nature is still disputed (Jolicoeur, Lefebvre, & Martinez-Trujillo, 2016). One interesting proposal ties in with predictive processing, that the icon consists of multiple layers, each of which is created by different neural processes and contains different sorts of information (Rensink, 2014; for a similar model, see Capiello & Zhang, 2016). For example, the first created and simplest layer is indeed an unprocessed representation of the physical stimulus that decays rapidly. Another layer is created after some conceptual processing, by neural circuits that feedback from higher areas of the brain to the input circuits of the retina, and it’s this layer that enables conceptual processing of sensory arrays as found by Merikle. On the other hand, Treisman & Lages (2013) argue that the icon and sensory memory simply do not exist as a stage of information processing. Using SDT, they propose that organisms adjust their decision criteria to optimally adapt to their sensory world and that, as decision criteria are adjusted, different aspects of sensory input, physical or semantic are observed.

研究模式识别

Investigating Pattern Recognition

因此，模式识别自认知科学诞生之初就受到广泛研究也就不足为奇了。同时，在认知科学的所有主要课题中，它始终处于理论边缘，一方面是心理学，另一方面是神经生理学的还原或消除。我们进行模式识别是如此无意识，以至于很难构建一个直观的、第一人称的描述。事实上，只有一种纯粹的心理学模式识别理论，而且它一开始就失败了。

It should not be surprising then that pattern recognition has been intensely studied from the first days of cognitive science. At the same time, of all the major topics in cognitive science, it has always been at the theoretical edge between psychology, on one side, and neurophysiological reduction or elimination, on the other. We do pattern recognition so unconsciously that it’s hard to formulate an intuitive, first-person account of it. Indeed, there’s only been one purely psychological theory of pattern recognition, and it failed out of the gate.

思想之路的失败

Failure of the Way of Ideas

联想主义的乐高意识理论在整个早期心理学中都具有影响力，但对其最执着的是一位在康奈尔大学度过职业生涯的英国心理学家——爱德华·布拉德福德·铁钦纳（Edward Bradford Titchener，1867-1927；铁钦纳，1897，1898；本费尔特，2013）。该理论也可以被称为意识的马赛克理论，认为诸如“狗”或“树”之类的复杂概念是由一些纯粹感觉的小片段构成的，但它们可以重新排列成“狗”或“树”或其他任何概念。

The associationist Lego theory of consciousness was influential in all early psychology, but it was pursued most tenaciously by an English psychologist who spent his career at Cornell, Edward Bradford Titchener (1867–1927; Titchener, 1897, 1898; Beenfeldt, 2013). It can also be called the mosaic theory of consciousness, holding that complex ideas, such as “dog” or “tree,” are made up of small bits of pure sensations, but they could be rearranged to be the ideas of “dog” or “tree,” or anything else.

在训练观察者时，铁钦纳（1912）严厉批评了未经训练的内省者和在竞争对手实验室接受训练的内省者所犯的“刺激错误”（铁钦纳，1912；博林，1921；哈特菲尔德，2015；奇里穆塔，2016）。如果我问你在图5.12中体验到了什么，你可能会说“一只狗”，犯了刺激错误，即说出了世界上的物理刺激，而不是构成你对它的心理体验的感觉。正如铁钦纳最重要的学生E.G.博林（1912）所说，你需要报告给定的，即模式识别随后运作的感官输入。随着内省分析在心理学中的进展，感官原子的列表将被汇编起来，就像物理学家和化学家汇编原子和亚原子的列表一样。元素。然后，联想连接将元素组合成诸如狗、猫、树、云等配置。

In training observers, Titchener (1912) railed against “the stimulus error” made by untrained introspectors and introspectors trained in rival laboratories (Titchener, 1912; Boring, 1921; Hatfield, 2015; Chirimuuta, 2016). If I ask you what you experience in Figure 5.12, you will probably say, “a dog,” committing the stimulus error, i.e., naming the physical stimulus in the world rather than the sensations comprising your mental experience of it. As Titchener’s most important student, E. G. Boring (1912) put it, you are to report the given, the sensory input upon which pattern recognition then operates. As introspective analysis proceeded in psychology, a list of sensory atoms would be compiled, just as physicists and chemists were compiling a list of atomic and sub-atomic elements. Associative connections then combined the elements into configurations such as dog, cat, tree, cloud, etc.

然而，对其他人来说，铁钦纳的贡献并非显而易见。威廉·詹姆斯在《心理学原理》中写道：

However, to others, it wasn’t obvious that Titchener’s given was theirs. In Principles of Psychology, William James wrote:

意识……本身并非被分割成碎片。“链条”或“火车”之类的词语并不能恰当地描述它，因为它最初呈现的样子（即，它是既定的）。它并非连接在一起的，而是流动的。“河流”或“溪流”是最自然的比喻。以后再谈论它时，我们不妨称之为思想之流、意识之流或主观生命之流。

James，1890 年，第 1 卷，第 239 页

Consciousness...… does not appear to itself chopped up in bits. Such words as “chain” or “train” do not describe it fitly, as it presents itself in the first instance [i.e., it’s the given]. It is nothing jointed; it flows. A “river” or a “stream” are the metaphors by which it is most naturally described. In talking of it hereafter let us call it the stream of thought, of consciousness, or of subjective life.

James, 1890, Vol. 1, p. 239

在德国，克里斯蒂安·冯·埃伦费尔斯（Christian von Ehrenfels，1859-1932）用音乐实例来推翻联想原子论。乍一看，一段旋律似乎由片段、单个音符组成，正如马赛克理论所言。然而，一段曲调可以调换成不同的调，其中所有音符都与之前的音符不同，但曲调依然存在。从某种意义上说，一段旋律是由一组音符加上埃伦费尔斯所谓的“形式品质”（Gestaltqualität）构成的，即一组音符之间的关系，这些关系无法被简化为一个感知原子。

In Germany, Christian von Ehrenfels (1859–1932) used musical examples to undermine the associative-atomistic theory of the given. At first glance, a melody appears to be made up of pieces, individual notes, just as mosaic theory says. However, a tune can be transposed into a different key in which none of the notes is the same as before, yet the tune remains. In some way, a melody is comprised of a set of notes plus what Ehrenfels called a form-quality (Gestaltqualität), a set of relations between the notes that cannot be reduced to a sensory atom.

格式塔心理学

Gestalt Psychology

格式塔心理学家给心理学中的原子论带来了致命一击，他们也开创了用生理学来解释模式识别的先河。格式塔心理学的主要代表人物是马克斯·韦特海默 (1880-1943)、沃尔夫冈·科勒 (1887-1967) 和库尔特·科夫卡 (1886-1941)。他们像艾伦费尔斯一样——他们的运动借用了艾伦费尔斯的名字——格式塔心理学——属于一个新的心理学哲学流派，即弗朗茨·布伦塔诺 (1838-1917) 的现实主义行为心理学，该流派成为具身认知的源泉之一。布伦塔诺认为，感知是一种将物体作为一个整体来把握的心理行为，而不是像原子论所教导的那样，将感觉原子记录下来，然后在心理上组装成更大的物体。

The death-blow to atomism in psychology was dealt by the Gestalt psychologists, who also inaugurated the turn to physiology in explaining pattern recognition. The leading Gestalt psychologists were Max Wertheimer (1880–1943), Wolfgang Köhler (1887–1967), and Kurt Koffka (1886–1941). Like Ehrenfels, from whom they borrowed the name of their movement, Gestalt psychology, they were members of a new philosophical school of psychology, the realist act psychology of Franz Brentano (1838–1917), which became one of the wellsprings of embodied cognition. Brentano contended that perception was a mental act of grasping an object as a whole, rather than the registration of sensory atoms that were then mentally assembled into larger objects, as atomism taught.

格式塔心理学家继承了布伦塔诺的观点，并将其发展成对“理念之路”的猛烈批判。他们的批判很好地展现了关于心灵的假设如何不假思索地从哲学转移到心理学，以及图表和程序的陷阱。回过头来看看笛卡尔的大脑感知箭的过程。它悄然体现了格式塔心理学家认为导致其科学研究脱轨的两个关于感知的假设。第一个是束假说（Wertheimer，1922），另一个是恒常性假说（Köhler，1947）。后者假设松果体上的每个感觉点都是由箭上特定点的刺激引起的。前者认为，我们对箭的最初感知——即给定——只是这些感觉点的束。换句话说，铁钦纳所谓的给定根本不是给定的，而是一个关于感知如何运作的哲学假设。铁钦纳心理学是为了寻找数据而寻找解释，而不是为了寻找解释而寻找数据。后来，科勒（1925）将同样的推理应用于桑代克未能在其谜题盒中找到推理的现象。被囚禁的猫无法看到盒子的运作方式，因此它们无法思考逃脱的方法。相反，它们将反复试验作为应对困境的唯一可能策略。实验装置，而非猫缺乏认知能力，导致了它们的行为。格式塔心理学家始终认为，科学必须从尽可能少的先入之见出发。科学的首要任务是描述自然，而不是将其强加于人类的概念框架中。从经验上讲，格式塔心理学始于一项实验，其中我们所体验到的事物没有刺激性的物理原因，这违反了捆绑假设和恒常性假设。韦特海默称之为^phi现象。5

The Gestalt psychologists took Brentano’s standpoint and developed it into a devastating critique of the Way of Ideas. Their criticisms are good examples of how assumptions about the mind can be unthinkingly carried over from philosophy to psychology, and of the traps of diagram and procedure. Go back and look at Descartes’s brain perceiving an arrow. It quietly embodied two assumptions about perception the Gestalt psychologists said had derailed its scientific study. The first was the bundle hypothesis (Wertheimer, 1922) and the other was the constancy hypothesis (Köhler, 1947). The latter assumes that every sensory point on the pineal gland was caused by stimulation from a particular point on the arrow. The former was that our initial perception of the arrow—the given—was just a bundle of these sensory points. In other words, Titchener’s so-called given was not given at all, but a philosophical assumption about how perception must work. Titchenerian psychology was explanation in search of data, not data in search of explanation. Later, Köhler (1925) applied the same line of reasoning to Thorndike’s failure to find reasoning in his puzzle boxes. The imprisoned cats could not see the workings of the box, and so they could not think their way to solutions for escape. Instead, they resorted to trial and error as the only possible strategy for dealing with their situation. The experimental setup, not cats’ lack of cognitive abilities, caused their behavior. Science, Gestalt psychologists always believed, had to start with as few preconceptions as possible. Science’s job is first to describe nature, not force it into human conceptual boxes. Empirically, Gestalt psychology began with an experiment in which what we experience has no stimulating physical cause, violating the assumptions of bundle and constancy. Wertheimer called it the phi phenomenon.⁵

客观上，只有两个球会亮起，然后随着另一个球的亮起而熄灭。然而，大多数观看者（存在个体差异）会感觉到左右和左右交替的运动。在理想情况下，观看者只会看到一个球在平滑连续地运动。我们在观看电影或视频时都会体验到phi现象，这些视频由一系列以每秒约24帧的速度播放的静止图像组成。如果你玩电子游戏时遇到过视频处理器变慢的情况，你就会发现 phi 现象的魔力正在令人沮丧地消失。因为你看到的是根本不存在的东西——一个与运动相对应的物理刺激点——所以 phi 现象现在被称为视运动。Wertheimer 武断地将其命名为phi，正是因为他不想让一个所谓的描述性标签误导他的参与者（Köhler 和 Koffka）的观察，或者他自己的理论。

Objectively, there are just two balls that light up and then turn off as the other ball is lit. However, most viewers (there are individual differences) sense motion alternating left-right and right-left. Under ideal conditions, viewers see only one ball in smooth, continuous motion. We all experience the phi phenomenon when we watch a movie or video, which are made of a series of still images shown at about 24 frames per second. If you play video games and ever have a slowdown in your video processor, you’ve seen the magic of the phi phenomenon frustratingly disappear. Because you see something that isn’t there—a point of physical stimulation to which the motion corresponds—the phi phenomenon is now called apparent motion. Wertheimer arbitrarily dubbed it phi precisely because he didn’t want an allegedly descriptive label to lead his participants’ (Köhler and Koffka) observations, or his own theorizing, astray.

我们无需运动就能看到感知如何违背原子论假设的另一个例子。图5.13展示了1955年发明的卡尼萨三角形，它显示了明显的轮廓。

We don’t need motion to see another example of how perception violates atomistic assumptions. Figure 5.13 shows the Kanizsa triangle, invented in 1955, showing apparent contours.

在这个图形中，大多数人看到的是一个倒置的等腰三角形，严格来说，它并不存在，只是一种暗示。此外，这个“幽灵三角形”往往看起来比背景更亮一些，在三角形拐角处，类似吃豆人图案的人物附近产生了虚幻的轮廓。与phi现象一样，这些轮廓的形成并不存在任何物理刺激点，这违反了粒子束和恒常性假说中的原子论假设。

In this figure, most people see an inverted isosceles triangle that, strictly speaking, is not there but only suggested. Moreover, this ghost triangle tends to look somewhat brighter than the background, generating illusory contours near the Pac-Man–type figures at the triangle’s corners. As with the phi phenomenon, there are no points of physical stimulation to cause the contours, violating the atomistic assumptions of the bundle and constancy hypotheses.

最后，格式塔心理学家可以指出，与马赛克石头不同，视觉显示的各个元素可以动态地相互作用，创造出超越其本身的体验。例如，图5.14是一个备受研究的错觉，它是由记忆实验研究的先驱赫尔曼·艾宾浩斯发现的。

Finally, Gestalt psychologists could point out that, unlike mosaic stones, the elements of a visual display may dynamically interact, creating experiences that go beyond them. In Figure 5.14, for example, is a much-studied illusion discovered by Hermann Ebbinghaus, the pioneer of the experimental study of memory.

圆圈中心的两个圆盘实际上是等大的。如果感知仅仅涉及看到感觉原子，那么就不会产生幻觉。

The two disks at the center of the circles are really the same size. If perception simply involved seeing sensory atoms, there would be no illusion.

总体而言，韦特海默（Wertheimer，1923/1938，第78页）与原子论相反，他指出：“当我们面对大量刺激时，我们通常不会体验到‘大量’个体事物，这一个，那一个……相反，经验中呈现出的是彼此分离又相互关联的更大整体……这种排列和划分是否遵循明确的原则？” 他的答案是肯定的。格式塔心理学家提出，大脑中存在一套与生俱来的感知法则，将感知组织成​​有意义的整体。例如，其中一条法则是闭合律，即我们倾向于在脑海中完成一个完整的图形。卡尼萨三角形就是闭合律发挥作用的一个例子：我们用明显的轮廓线填充缺失三角形的线条。后来，科夫卡（Koffka，1935/1963）将闭合律阐述为更普遍的格式塔原理，即普朗格纳兹法则（Prägnanz Law） ，即简单性法则。我们自然而然地将经验分组为可用的最少数量的简单图形。

In general, Wertheimer (1923/1938, p. 78) said, contrary to atomism, “When we are presented with a number of stimuli we do not as a rule experience ‘a number’ of individual things, this one and that…. Instead larger wholes separated from and related to one another are given in experience… Do such arrangements and divisions follow definite principles?” He answered yes, and the Gestalt psychologists proposed that there was a small set of innate principles of laws of perception built into the brain that organize percepts into meaningful wholes. For example, one of the principles, is the law of closure, our tendency to mentally complete a good figure. The Kanizsa triangle is an example of closure at work: we fill in the lines of the missing triangle with apparent contours. Later, Koffka (1935/1963) elaborated the law of closure into a more general Gestalt principle, the Law of Prägnanz, or simplicity. We naturally group experiences into the smallest number of simple figures available.

格式塔原则的有效性已在平面设计中得到广泛应用。6^例如，这里有一个你见过很多次的著名商业标志，但你可能没有注意到我所说的“蹲伏格式塔，隐藏的箭头”：

That Gestalt principles work is testified to their extensive use in graphic design.⁶ For example, here’s a famous commercial logo you’ve seen many times, but you might not have noticed what I call Crouching Gestalt, Hidden Arrow:

直到我读到一篇关于这个标志如何诞生的文章时，我才注意到（当第一次看到这个标志时，大多数联邦快递董事会成员都没有注意到）Ex 中的箭头；当然，现在我无法忽视它了。

Until I read an article about how the logo came to be I did not notice (as most of the FedEx board didn’t when they were first shown it) the arrow in the Ex; now, of course, I can’t unsee it.

格式塔心理学家有效地推翻了乐高模式识别理论。然而，他们取而代之的理论并非心理学理论。他们与巴甫洛夫共享脑场理论，并提出脑电场相互作用能够反映或映射物理宇宙的结构。正如科勒（Köhler，1967/1971，第115页）所写：“格式塔心理学自此成为场物理学在心理学和脑生理学重要领域的一种应用。” 即使认知科学进入计算时代，这种转向神经科学的趋势依然存在。

The Gestalt psychologists effectively destroyed the Lego theory of pattern recognition. However, the theory they put in its place was not psychological. They shared with Pavlov the field theory of the brain and proposed that the electric field interactions of the brain came to reflect, or map, the structure of the physical universe. As Köhler (1967/1971, p. 115) wrote, “Gestalt psychology has since become a kind of application of field physics to essential parts of psychology and brain physiology.” The turn to neuroscience remained even when cognitive science entered the computational era.

神经网络和神经科学

Neural Networks and Neuroscience

由于各种原因（Leahey，2018），格式塔心理学对原子论的批判影响了后来的认知心理学，但他们的实证理论却没有。7^就模式识别而言，他们的原则被认为依赖于过时的神经生理学理论，更重要的是，它们只是描述性的，而非解释性的。也就是说，格式塔原则优雅地捕捉了经验在我们看来的样子，但并没有真正一旦人们不再将大脑视为一组电磁场的观点，就可以解释它们的工作原理。

For a variety of reasons (Leahey, 2018), Gestalt psychology’s critique of atomism influenced subsequent cognitive psychology, but their positive theories did not.⁷ In the case of pattern recognition their principles were seen as depending on an outmoded theory of neurophysiology, and more importantly as being descriptive but not explanatory. That is, Gestalt principles elegantly captured how experience seems to us, but did not really give an account of how they worked, once one discarded the view of the brain as a set of electromagnetic fields.

特征理论：混乱

Feature Theory: Pandemonium

第一个模式识别的计算理论是由 Oliver Selfridge (1959; Selfridge & Neisser, 1960) 提出的。它被称为Pandemonium，和联结主义一样，它由多层简单计算单元组成，这些计算单元可以通过改变它们之间的链接权重来学习。在弥尔顿的《失乐园》中，上帝打败了撒旦和他的反叛天使后，他们被投入地狱，在那里他们建造了一个他们共同居住的地方——Pan ( all) demonium。塞尔福里奇的理论是认知科学和人工智能中一般策略的一个很好的例子，它将一个复杂的任务（在本例中为模式识别）分解为一组更简单的任务，这些任务似乎需要复杂的智能，而这些任务可以由没有思想的机器人或神经元来完成。在 Pandemonium 中，机器人是恶魔，每个机器人都有一项简单的、无需思考的工作要做。尽管如此，只要积累足够多的没有思想的恶魔，智慧就会出现（Dennett, 1983）。

The first computational theory of pattern recognition was proposed by Oliver Selfridge (1959; Selfridge & Neisser, 1960). It was called Pandemonium, and, as in connectionism, it consisted of multiple layers of simple computing units that could learn by changing the weights of the links between them. In Milton’s Paradise Lost, after God defeats Satan and his rebel angels, they are cast into Hell where they build a place where they all live—Pan (all) demonium. Selfridge’s theory is an excellent example of a general strategy in cognitive science and AI, breaking a complex task, in this case pattern recognition, that appears to demand sophisticated intelligence into a set of much simpler tasks that can be carried out by thoughtless bots or neurons. In Pandemonium the bots are demons each of which has one, simple, mindless job to do. Nevertheless, add up enough thoughtless demons and intelligence emerges (Dennett, 1983).

Pandemonium 的总体策略是特征分析。正如我们之前所讨论的，在模式识别中，对世界的呈现必须与存储在记忆中的概念相匹配。理解概念的一种方法是，它们包含一系列定义特征，就像字典一样。例如，“单身汉”指的是未婚男性。如果我们遇到拥有这些属性或特征的人，那么这个人就是单身汉。所以，所有模式识别的工作原理可能都是一样的。Pandemonium 最常见的例子是识别字母，这是人工智能的首要目标之一。考虑到当今先进的文本扫描和翻译技术，这似乎并非难题，但计算机花了几十年的时间才能够扫描书籍并创建文本文件。

The general strategy of Pandemonium is feature analysis. In pattern recognition, as we’ve discussed, a presentation of the world must be matched up with a concept stored in memory. One way to think about concepts is that they comprise a list of defining features, as in a dictionary. So, for example, a bachelor is an unmarried male. If we meet someone possessing these attributes, or features, that person is a bachelor. So it could be that all pattern recognition works the same way. The most common example of Pandemonium at work was recognizing letters, one of the first goals of AI. Given today’s sophisticated technology for scanning and translating text, it doesn’t seem like a hard problem, but it took decades for computers to be able to scan books and create text files.

最早的程序使用一种叫做模板匹配的方法。计算机内存中存储着字母的小图像（模板）。A，a；A，a；A，a；等等。然后，扫描仪扫描文本，模板匹配程序会确定哪个字母（如果有的话）与扫描字母的物理形状匹配。然而，正如我在上文中部署的几种字体所示，即使是字体大小、形状、扫描平台的角度、墨水渗入廉价纸张等方面的细微偏差，都可能使程序崩溃，导致输出乱码。Pandemonium 提供了一种更强大的替代方案，而如今它的深度学习后代只是使用了更复杂的策略版本（图 5.15）。

The earliest programs used a method called template matching. In computer memory were stored little images, templates, of letters. A, a; A, a; A, a; and so on. A scanner then scanned text, and the template matching program determined what, if any, letter matched the physical shape of the scanned letter. However, as the few fonts I deployed in the sentence above show, even small deviations of font size, shape, angle on the scanning platform, ink bleeding through cheap paper, etc. could defeat the program, resulting in gibberish. Pandemonium offered a more robust alternative, and its deep learning descendants today simply use more sophisticated versions of its strategy (Figure 5.15).

我们从图像恶魔开始，它基本上就是斯珀林的图标，它手持输入字母的呈现，等待第二层恶魔——特征恶魔——的处理。它们会检查图标上指定的图形特征，例如垂直线、水平线或曲线，并根据图标上显示的特征数量（塞尔福里奇语）发出尖叫声。第三层由认知恶魔组成，每个认知恶魔代表一个由其关键特征定义的字母表字母，它们尖叫的程度取决于它的许多定义特征都已存在。因此，R 被定义为一条垂直线、两条水平线、一条斜线、三个直角和一条不连续的曲线。但这些特征也出现在其他字母（例如 P）的定义中，因此多个恶魔同时尖叫。因此，需要再加一层来完成模式识别的过程。最后一层只有一个恶魔，即决策恶魔，但它的工作与其他恶魔一样简单，就是从之前的尖叫声中挑选出最响亮的那个。恶魔层。在这种情况下，虽然 P 恶魔和 R 恶魔都在大声尖叫，但声音最大的是 R 恶魔，而决策恶魔将输入模式识别为R。

We begin with the image demon, basically Sperling’s icon, who holds a presentation of the input letter for processing by the second layer of demons, the feature demons. They inspect the icon for their assigned graphic feature, such as a vertical line, horizontal line, or curve, shrieking (Selfridge’s word) in proportion to the number of its features the icon displays. The third layer consists of the cognitive demons, each of whom represents a letter of the alphabet defined by its critical features, and it shrieks as a function of how many of its defining features are present. So, an R is defined as one vertical line, two horizontal lines, an oblique line, three right angles, and one discontinuous curve. But these features also figure in the definition of other letters, such as P, and multiple demons are therefore shrieking in parallel. One more layer is therefore needed to complete the process of pattern recognition. This final layer has just one demon, the decision demon, but its job is just as simple as the others’ jobs, to pick the loudest shrieker from the earlier layer of demons. In this case, while the P and R demons are shrieking loudly, the loudest is the R demon, and the decision demon identifies the input pattern as an R.

在许多方面，Pandemonium 堪称多层前馈人工神经网络（ANN）联结主义架构的先驱。然而，它面临两个重要的局限性。首先，误差信号反向传播的技术尚未开发。从概念上讲，Pandemonium 仍然依赖于符号系统架构，在这种架构中，系统启动时必须将一定程度的知识编程到系统中。特征和认知“恶魔”已为过程分配了符号，特征“恶魔”为直线、曲线和角度，认知“恶魔”为有意义的字母。因此，Pandemonium 与其说是发现了世界中的模式，自己学习了什么是垂直线或什么是R，不如说是学会了将预先编程的知识应用于具体案例。

In many respects, Pandemonium was a pioneer of connectionist architectures of multilayered feedforward ANNs. However, it faced two important limitations. First, the technique of backpropagation of error signals had not yet been developed. More conceptually, Pandemonium was still tied to the symbol system architecture in which some degree of knowledge had to be programmed into the system at the beginning. The feature and cognitive demons had assigned symbols to process, lines, curves, and angles in the case of the feature demons, and meaningful letters in the case of the cognitive demons. Thus, Pandemonium did not so much discover patterns in the world, learning on its own what a vertical line is or what an R is, as it learned to apply its pre-programmed knowledge to specific cases.

调查注意力

Investigating Attention

在本章中，我们一直使用序列处理模型来佐证注意力，该模型认为注意力是信息处理过程中一个独特的阶段或时刻，詹姆斯正是在这个阶段“掌控”了一系列的思路。如果我们假设这一点，就会出现两个经验问题。首先，注意力选择发生在哪个阶段？是早期，接近图像，还是晚期，就在意识之门之前？第二个问题与第一个问题相关，选择的依据是什么？如果选择发生在早期，在图像和模式识别之间，那么选择似乎取决于刺激的物理特征，例如它在我们的右侧还是左侧，是音乐而不是对话，是视觉刺激而不是听觉刺激。如果选择发生在晚期，发生在模式识别之后——刺激变得有意义的时候——但在意识形成之前，那么我们可以根据所见所闻的意义来选择，同意关注我们会接受令我们高兴的事情，并拒绝接受令我们不悦的事情。然而，图5.16中漫画的幽默暗示着情况并非如此简单。两个塞尔福里奇式的恶魔正在监听一位体育迷的音频输入。请注意，我们是在球迷的耳朵里，所以处理过程还处于早期阶段。然而，左边的恶魔正在根据配偶的话语内容“占据”球迷的大脑，这表明模式识别已经发生。

In the serial processing model of attention we have been using as our foil in this chapter, attention is thought to be a distinct stage, or moment, in information processing at which James “taking possession” of a train of thought takes place. If we assume this, then two empirical questions arise. First, at what point does attentional selection take place? Is it early, near the icon, or late, just before the doorway to consciousness? The second question is related to the first, on what basis is selection made? If the choice is early, between the icon and pattern recognition, the choice would seem to have to depend on the physical features of the stimulus, such as its being to our right or left or being music rather than talk or visual rather than auditory. If the choice is late, occurring after pattern recognition—when stimuli become meaningful—but before consciousness, then we could choose on the basis of the meaning of what we see or hear, agreeing to attend to things that please us and refusing to agree to things that displease us. The humor in a cartoon such as that in Figure 5.16 suggests, however, that the picture is not so simple. Two Selfridge-like demons monitor a sports fan’s audio input. Notice that we’re in the fan’s ear, so processing is early. Nevertheless, the demon on the left is “taking possession” of stimuli for the fan’s mind on the basis of the content of what the spouse says, suggesting pattern recognition has already happened.

有趣的是，最早的现代注意力实验研究是针对听觉而非视觉注意力进行的，就像动画片中的设定一样。这在一定程度上是商业对认知科学的影响，就像今天的人工智能一样，因为电话公司（没错，曾经只有一家电话公司，而且他们拥有电话）对优化通过简单的铜线和廉价电话进行的语音传输非常感兴趣。像E·科林·切里（E. Colin Cherry，1953）这样的工程师开发了双耳分听程序，以模拟现实生活中一种被称为鸡尾酒会效应的选择性注意案例。

Interestingly, the first modern experimental studies of attention were carried out for auditory rather than visual attention, like the setting in the cartoon. In part this was an influence of commerce on cognitive science, as in AI today, because the telephone company (yes, there was just one once, and they owned the phones) had great interest in optimizing voice transmission over simple copper wires and cheap telephones. Engineers such as E. Colin Cherry (1953) developed the dichotic listening procedure to mimic a real-life case of selective attention called the cocktail party effect.

如果你身处一个热闹的鸡尾酒会，你必须在周围嘈杂的谈话声和噪音中，用心聆听朋友的谈话。你听到朋友说了什么，其他人的谈话则沦为背景中舞台演员的“喋喋不休，喋喋不休；咕噜咕噜”，毫无意义的声音。在双耳分听，参与者佩戴耳机，注意力集中在一只耳朵的输入上。通常情况下，参与者会被要求在被关注的耳朵里跟读材料，尽可能准确地重复听到的内容。所以，如果你关注左耳并听到“在人类事件的过程中……”，你必须重复你听到的每个单词。这是一项要求很高的任务，通常大约 20% 的参与者会因为无法跟读而退出实验。要求较低的注意力任务包括从提供的列表中数出名词、押韵词或同义词。其他类型的输入可以在未被关注的耳朵中播放，最后，我们可以测试参与者对双耳所听内容的记忆。

If you are at a busy cocktail party, you must mentally take possession of a friend’s conversation among the cacophony of conversations and noises going on around you. You hear what your friend says, other talk being reduced to what stage actors in the background say, “natter, natter; grommish, grommish,” vocal sounds sans meanings. In dichotic listening, the participant wears headphones, and attention is directed to one ear’s input. Typically, the participant is asked to shadow the material in the attended ear, repeating it as closely as possible as it’s heard. So, if you attend the left ear and hear “When in the course of human events…,” you must repeat each word as you hear it. This is a demanding task, and often about 20% of participants wash out of the experiment for not being able to shadow. Less demanding attention tasks include counting nouns or rhymes or synonyms from a presented list. Input of various other types can be played in the unattended ear, and, at the end, we can test the participant on what they remember about what they heard in both ears.

双耳分听研究的基本发现（除了一个我们将要讨论的异常现象）与我们根据鸡尾酒会的亲身经历以及直觉序列信息处理模型所预期的结果相符。参与者可以概括被关注通道的要点，但无法概括未被关注通道的要点，但他们可以回忆起未被关注通道的物理特征，例如男性和女性的声音，或者无关刺激（例如音调）的介入。最早的注意力信息处理模型就是围绕这一范式及其结果建立的。

The basic findings (with one anomaly we’ll come to) from dichotic listening studies replicated what we would expect from our personal experience of cocktail parties and with the intuitive serial information processing model. Participants could summarize the gist of the attended, but not the unattended, channel, but they could recall physical characteristics of the unattended channel, such as male vs. female voice, or the intrusion of an irrelevant stimulus, such as a tone. The first information processing models of attention were built around this paradigm and its results.

注意力理论

Theories of Attention

除了他对注意力的著名描述之外，詹姆斯还对两类注意力理论——因果理论和因果理论——进行了有益的区分。因果理论认为，存在一个心理处理阶段，或者说，将注意力资源分配给一个输入流，以便（正如詹姆斯所说）有效地处理体验的某一方面。在上面的漫画中，操作风扇耳朵的恶魔体现了因果理论，因为它选择进入风扇意识的内容。另一方面，在意识的因果理论中，进入大脑的内容是由注意力以外的心理过程决定的。例如，在一部恐怖电影中，一个角色经常独自置身于可怕的环境中，哪怕是最轻微的声音都会让他惊慌失措。角色的广泛性焦虑导致原本微弱的刺激被放大并进入大脑。让詹姆斯感到苦恼的是，所有注意力，即使是看似自愿的注意力，都可能是心理原因而非选择的结果。詹姆斯的担心得到了证实，因为随着时间的推移，效果理论取代了原因理论，直到一些认知科学家声称注意力根本不存在（Anderson，2011）。

In addition to his famous description of attention, James made a useful distinction between two classes of theories of attention, cause theories and effect theories. Cause theories of attention posit the existence of a mental stage of processing or of allocation of attentional resources to one stream of input in order, as James put it, to deal effectively with just one aspect of experience. In the cartoon above, the demon operating the fan’s ear embodies a cause theory of attention because it chooses what passes into the fan’s consciousness. In an effect theory of consciousness, on the other hand, what enters the mind is determined by a mental process other than attention. For example, in a horror movie a character is often alone in frightening surroundings, and startles at the least little sound. The character’s generalized anxiety caused an otherwise weak stimulus to be magnified and admitted to the mind. What came to vex James was the possibility that all attention, even seemingly voluntary attention, was the outcome of mental causes rather than choice. James’s fear was borne out, as over time effect theories replaced cause theories until it was possible for some cognitive scientists to claim that there’s no such thing as attention (Anderson, 2011).

原因论

Cause Theories

现代第一个注意力理论是 Broadbent (1958) 的过滤理论，它也很好地说明了现有技术如何为注意力提供隐喻。大脑。布罗德本特提出，注意力的作用类似于收音机或电视上的频道选择器。许多广播到达设备的天线，但调谐器只选择其中一个进行处理（图5.17）。

The first theory of attention in the modern era was Broadbent’s (1958) filter theory, which also provides a nice example of how existing technology has provided metaphors for the mind. Broadbent proposed that attention work liked the channel selector on a radio or TV. Many broadcasts reach the device’s antenna, but your tuner selects just one to process (Figure 5.17).

注意是信息处理中一个独特的阶段，位于图标之后，因此它会根据图标上可用的输入刺激的物理特性进行选择，例如空间位置（例如，在聚会上你会面向朋友），或者与周围其他声音相比，被关注的声音的音调，并且你可以注意到新声音的进入。你可以总结被关注的耳朵听到的内容，因为它在语义上被处理并进入记忆，但你无法描述未被关注通道的内容，因为它没有被处理。简而言之，过滤理论解释了早期双耳分听实验的基本结果。

Attention is a distinct stage in information processing located just after the icon, and thus making its selection based on physical characteristics of the input stimulus available on the icon, such as location in space (you orient toward your friend at a party) or pitch of the attended voice compared to other voices near you, and you can note the intrusion of novel sounds. You can summarize what the attended ear heard, because it was semantically processed through to memory, but you can’t say anything about the content of the unattended channel because it wasn’t. In short, filter theory explains the basic results of early dichotic listening experiments.

然而，后来的研究表明，Broadbent 的过滤机制过于强大。Moray（1959）将参与者的名字放入无人注意的频道，发现它经常被注意到。在过滤账户中，这种情况不应该发生，因为名字不应该在图标之后被处理。Broadbent 的学生 Anne Treisman（1964、2004、2012；Treisman & Gelade，1980）发展了一种改进的过滤理论，以适应 Moray 和其他类似研究的结果（图 5.18）。

However, later findings showed that Broadbent’s filter was too strong. Moray (1959) dropped participants’ names into the unattended channel, discovering that frequently it was noticed. On the filter account this should not happen, because the name should have not been processed past the icon. Broadbent’s student Anne Treisman (1964, 2004, 2012; Treisman & Gelade, 1980) developed a modified filter theory to accommodate Moray’s results and others like them (Figure 5.18).

注意力在信息处理中仍是一个独特的阶段，但它不再是我们选择进一步处理内容的唯一决定因素。它不再选择一个思路并阻止其他思路，而是削弱未选择的输入强度。长期记忆现在在将思想纳入意识方面发挥着作用。特雷斯曼提出，存储在长期记忆中的想法，例如金丝雀、联结主义或你的名字，具有不同的激活阈值，需要一定量的输入能量才能被激活。因此，被关注通道中的想法往往会被激活并成为意识，因为它们的输入能量被允许以最大强度通过。大多数无人注意通道中的想法不会被激活，因为无人注意流的强度会降低到阈值以下。然而，与此同时，那些阈值较低的想法，比如你喜欢听到的“你的名字”，即使在输入强度较低的情况下也会被激活。所以，你会听到“喋喋不休，喋喋不休，咕噜咕噜，你的名字，咕噜咕噜，喋喋不休。”

Attention is still a distinct stage in information processing, but it is no longer the sole determiner of what we select for further processing. Instead of choosing one stream of thought and blocking others, it now weakens the input strength of unselected inputs. Long-term memory now plays a role in admitting thoughts to consciousness. Treisman proposed that ideas stored in long-term memory, such as canary, or connectionism, or your name, have different thresholds of activation, requiring a certain amount of input energy to become activated. Thus, ideas in the attended channel will tend to activate and become conscious because their input energy is allowed through at full strength. Most ideas in the unattended channel will not activate because the strength of the unattended stream is reduced to below their threshold values. At the same time, however, ideas that have low threshold values, such as your name, which is something you like to hear, will activate even at low input strengths. So, you hear, “Natter, natter, grommish, YOUR NAME, grommish, natter.”

特雷斯曼的理论摒弃了串行处理。由于无人参与的信息流仍然会被处理，尽管强度有所减弱，但它们最终会进入长期记忆，并可能对意识体验产生影响。特雷斯曼的理论是并行处理概念被发明之前的先驱模型。在特雷斯曼的模型中，注意力仍然是你体验的原因之一，但并非唯一原因。在哲学和科学领域，奥卡姆剃刀原则告诉我们，要用尽可能少的理论实体来解释事物。或许，注意力是一个我们可以省略的原因或处理阶段，这解释了直觉和詹姆斯所说的注意力是其他信息处理的结果。

Treisman’s theory abandoned serial processing. Because unattended streams are still processed, albeit at attenuated strength, they reach long-term memory, where they may have an effect on conscious experience. Hers was a pioneering model of parallel processing before the concept had quite been invented. In Treisman’s model, attention is still a cause of what you experience, but not the sole cause. In philosophy and science, Ockham’s razor tells us to explain things with as few theoretical entities as possible. Perhaps attention is a cause or stage of processing we can do without, explaining what intuition and James call attention as the result of other information processes.

效应理论

Effect Theories

第一个沿着这一思路提出的观点来自 Deutsch & Deutsch (1963, p.83)：“无论是否关注，信息都会到达相同的感知和辨别机制”（图 5.19）。

The first proposal along these lines came from Deutsch & Deutsch (1963, p. 83), “a message will reach the same perceptual and discriminatory mechanisms whether attention is payed to it or not” (Figure 5.19).

在这个理论中，信息处理中不存在注意力选择阶段。相反，所有输入都会被完全处理以产生意义，而你所听到内容的意义决定了其是否能够进入意识。Deutsch 和 Deutsch 对其运作方式含糊其辞。但他们确实提出了一个在认知科学中日益重要的问题，尤其是在新的预测处理框架中，即所谓的感知的认知渗透性。传统上，我们认为感知是一个自下而上的过程，通过我们的感官感知世界并对其进行表征——世界强加于我们——然后我们必须掌控我们所见所闻。然而，如果感知能够被我们已知的、希望的和恐惧的事物所渗透，那么世界就不是以经验的形式强加于我们，而是我们强加于自身。接下来的问题是如何以及在何种程度上强加于我们。

In this theory, there is no stage of attentional selection in information processing. Instead, all inputs are fully processed for meaning and it’s the meaning of what you hear that determines access to consciousness. Deutsch and Deutsch were carefully vague on how it works. But they did raise an issue of increasing importance in cognitive science, especially in the new predictive processing framework of what’s called the cognitive penetrability of perception. Traditionally, we think of perception as a bottom-up process by which our senses detect the world and represent it—the world forces itself upon us—and then we must take possession of what we see and hear. If, however, perception can be penetrated by what we already know, and hope, and fear, then the world does not force itself upon us as experience, we force ourselves on it. The question then becomes how and to what degree.

研究方法从听觉转向视觉，推动了后来的注意力理论朝着多伊奇和多伊奇的方向发展。乌尔里克·奈瑟（Ulric Neisser）的原著《认知心理学》（1967）为该领域的建立做出了巨大贡献，如今却成了该领域的批评者（Neisser，1976）。他的主要抱怨是认知科学过于退缩到实验室，脱离了现实世界。就注意力而言，他认为二分式的聆听范式尤其人为，围绕它建立的理论过于依赖听觉注意力，而视觉对人类更为重要，这使得过滤式理论显得荒谬。奈瑟发明了“隐形大猩猩”实验（Neisser & Becklen，1975）。

A shift in research methods from audition to vision pushed later theories of attention in Deutsch and Deutsch’s direction. Ulric Neisser, whose ur-text Cognitive Psychology (1967) had done so much to establish the field, now became its critic (Neisser, 1976). His main gripe was that cognitive science had retreated too much into the lab, getting out of touch with the real world. In the case of attention, he thought the dichotomous listening paradigm was especially artificial, and theories built around it were too tied to auditory attention, when vision was more important to humans and made filter-type theories absurd. Neisser invented the invisible gorilla experiment (Neisser & Becklen, 1975).

Neisser (1976) 非正式地表达了他的观点。想象一下，你在剧院观看芭蕾舞演出。你对一位即将表演特别高难度曲目的新任首席芭蕾舞女演员的才华很感兴趣。演出结束后，我问发生了什么事，你告诉我这位新任芭蕾舞女演员的情况和她的技巧，以及她是如何完成首秀的。我问舞台设计怎么样？芭蕾舞团（背景舞者）怎么样？你几乎无法告诉我什么，因为你关注的是这位新星的表演，而不是舞台布景或其他舞者。但是，舞台布景和芭蕾舞团并没有像过滤理论那样被遮挡在视野之外，也没有像特雷斯曼的理论那样变成被削弱的幽灵。你兴趣之外的信息并没有被拒绝；只是没有被捕捉到而已。如果你对舞台布景感兴趣（你最好的朋友，帮你买到票的人，是剧院的首席设计师），你就会捕捉到关于舞台布景的信息。如果你的孩子首次参加芭蕾舞团演出，你就会获得这些信息。滤波器和滤波器衰减理论只有在基于电话或无线电传输的双耳分听范式中才有意义。

Informally, Neisser (1976) put his message this way. Imagine you’re at a theater, watching a ballet. You are interested in the talents of a new prima ballerina who will be dancing an especially difficult piece. After the show, I ask what happened, and you tell me about the new ballerina and her skills, and how she pulled off her first appearance. I ask what the set design was like? How was the corps de ballet (background dancers)? You’ll be able to tell me very little because you attended to the new star’s performance, not the sets or the other dancers. But the sets and corps did not get blocked from view, as in filter theory, nor did they become attenuated ghosts, as in Treisman’s theory. Information outside your interests was not rejected; it simply did not get picked up. Had you been interested in the sets (your best friend, who got you the tickets, is the theater’s lead designer), you’d have picked up information about them. Had your child been making its debut in the corps de ballet, you’d have picked up that information. The filter and filter-attenuation theories only made sense in the dichotic listening paradigm, based on telephone or radio transmission.

奈瑟的观点很有吸引力，因为它尊重我们（以及詹姆斯）对注意力的直觉——注意力关乎我们选择处理什么——但它放弃了这样一种观念：在一系列允许某些输入、阻止其他输入的时刻中，存在一个单一的选择时刻。相反，信息处理沿着多条平行的轨道进行，因此，我们的兴趣可以在许多地方和时间塑造体验（Stigchel，2016）。

Neisser’s view is attractive because it respects our (and James’s) intuitions about attention—that it’s about what we choose to process—but it gives up the notion that’s there’s a singular moment of choice in a series of moments where some input is allowed in and the rest is blocked. Instead, information processing takes place along multiple parallel tracks, and, thus, there are many places and times our interests can shape experience (Stigchel, 2016).

认知盲点：专注的危险

Cognitive Blindness: Perils of Focus

在本章中，我们阐述了自上而下、概念驱动的感知和模式识别的影响。一旦认识到这一点，我们就有可能发现自上而下的期望和注意力集中可能导致我们错过周围发生的事件，从而造成所谓的认知盲区（因为这种盲区是由自上而下的信息处理引起的）。

In this chapter we have been demonstrating top-down, conceptually driven influences on perception and pattern recognition. Once recognized, it becomes possible to discover instances in which top-down expectations and attentional focus can cause us to miss events occurring around us, creating what can be called cognitive—because caused by top-down information processes—blindness.

先验从何而来？ 一、进化

Where Do Priors Come from? I. Evolution

本书以认知科学的哲学和认识论视角开篇。人类心智的目标是认识真理，或者更心理学意义上的认识论，即获得可证实的真实信念。自苏格拉底以来，整个认识论体系，除了偶尔受到怀疑论者的抨击外，都致力于揭示心智如何能够并且确实认识真理。福多尔（2000）坚持这条道路，甚至拒绝了进化论和进化心理学。然而，预测处理架构接受进化论，因此，它将人类心智的最终目标是实现对世界的有效行为适应。真理固然美好，但生存更重要，而福多尔的坚持实际上是一种狂热（Frankenhuis & Ploeger，2007；Pinker，2005）。

We began this book with the philosophical, epistemological perspective on cognitive science. The goal of the human mind is to know Truth or, more psychologically, to acquire justifiable true beliefs. The entire enterprise of epistemology from Socrates onward, with the exception of occasional sniping by skeptics, was devoted to showing how the mind could, and did, know the truth. Fodor (2000) tenaciously kept to this path to the point of rejecting evolution and evolutionary psychology. However, the predictive processing architecture accepts evolution, and, therefore, it sees the human mind’s ultimate goal as achieving effective behavioral adaptation to the world. Truth is nice, but survival is more important, and Fodor’s tenacity is really fanaticism (Frankenhuis & Ploeger, 2007; Pinker, 2005).

因此，进化很可能塑造了感知，使其能够最佳地应对——拥有有效的贝叶斯先验——心智将要置身于的特定世界。我们天生就具备随时可用且可靠适应的意识。例如，请考虑一下Kleffner和Ramachandran（1992）的弹出点（图5.20）。

Therefore, it is likely that evolution has shaped perception to be optimally prepared for—have effective Bayesian priors for—the particular world in which the mind will find itself. We come equipped with a consciousness ready to hand and reliably adaptive. Consider, for example, Kleffner and Ramachandran’s (1992) pop-out dots (Figure 5.20).

这种错觉以及其他错觉的产生都是因为，包括人类在内的动物都是在一个由上方单一光点（太阳）照亮的世界中进化而来的（Stone, Kerrigan, & Porrill, 2009; Gigerenzer, 2008; Clark, 2013, 2016）。

This illusion and others are caused by the fact that animals, including people, have evolved in a world illuminated by a single point of light (the sun) shining from above (Stone, Kerrigan, & Porrill, 2009; Gigerenzer, 2008; Clark, 2013, 2016).

这种先天先验的威力或许可以用一种名为“反遮蔽”的现象来最好地体现。这种现象被动物（Tinbergen，1957）和人类用作伪装，是由美国艺术家阿博特·塞耶（Abbot Thayer，1849-1941；Behrens，2014）发现的。当我们看到凸出的点时，我们的视觉系统会将中间两列看作是朝我们凸出的半球：假设这些点都是均匀的灰色，那么它们在光线照射到的顶部应该呈现较浅的灰色，而在阴影下的底部应该呈现较暗的灰色。当我们翻转圆点，两个半球现在会从我们眼前弹出，光线到达底部的地方颜色较浅，而顶部则较暗。塞耶推断，如果一个物体被涂成顶部比底部更暗（仿佛在阴影中），就像在光照下一样，我们的视觉就会被这种刻意的反阴影所欺骗（图5.21）。

The power of this innate prior is perhaps best illustrated by a phenomenon called countershading that is used as camouflage by both animals (Tinbergen, 1957) and humans, discovered by American artist Abbot Thayer (1849–1941; Behrens, 2014). In seeing the pop-out dots, our visual system sees the middle two columns as hemispheres popping out toward us: Assuming that the dots are uniformly grey, they should be lighter grey on top where light strikes them and darker below, where there is shade. When we flip the dots, the hemispheres now are taken to pop away from us, paler where the light reaches the bottom and darker on top. Thayer reasoned that if an object were painted such that its top was darker (as if it were in shade) than its bottom (as if it were lighted), our vision would be fooled by this deliberate counter-shading (Figure 5.21).

反遮光技术在进化过程中已经被使用，经过塞耶和他的儿子的长期努力，它成为了军事伪装的主要手段（图 5.22）。

Countershading was already used by evolution, and, after a long campaign by Thayer and his son, it became a staple of military camouflage (Figure 5.22).

我们大脑关于地球上物体如何被照亮的固有假设的最后一个例子是颜色恒常性现象。乔治·贝克莱指出了尺寸恒常性现象，即尽管物体在我们视网膜上的投射尺寸发生了变化，但我们对物体尺寸的感知保持不变。我们对颜色也有同样的感知，尽管物体的实际亮度随着从亮处移到暗处而发生变化，我们对物体颜色的感知保持不变^（图5.23）。

A final example of our brain’s innate assumption about how objects on earth are lit is the phenomenon of color constancy. George Berkeley, recall, pointed out the phenomenon of size constancy, how we hold constant in perception an object’s size despite changes in its projected size on our retinas. We do the same with colors, holding constant in perception an object’s color despite changes in its actual illumination as it moves from light into shade¹¹ (Figure 5.23).

总的来说，这些现象表明，刺激误差无需克服，因为在人类信息处理的任何环节，都不存在完美的感觉——未经解读的简单原子刺激。此外，我们的许多感知先验早在人类出现之前就已进化，至少存在于所有哺乳动物的大脑中。

Collectively, these phenomena suggest that there is no stimulus error to be overcome because there are no immaculate sensations—uninterpreted simple atomic stimuli—existing at any point in human information processing. Moreover, many of our perceptual priors evolved well before the appearance of humans, existing in at least all mammalian brains.

先验从何而来？II. 概念

Where Do Priors Come from? II. Concepts

与动物不同，人类会学习概念，而这些概念也能渗透到感知中（Lupyan，2017）。像贾斯特罗的鸭子/兔子这样的模糊图形，通常一次只能被看作一种解读，要么是鸭子，要么是兔子，并在瞬间从一种解读转换为另一种解读（Lupyan，2017）。然而，在一个巧妙的实验中，詹森和马修森（2011）设计了一些言语提示，让受试者在两种解读下都能看到相同的模糊图形对。图 5.24展示了鸭子/兔子对。参与者最初看到这些图像时，通常都会以一种模棱两可的方式看待它们，他们的体验会从一种解读转换到另一种解读。然而，当听到“鸭子正在吃兔子”这个提示时，大多数人都能将右侧图像看作是捕食者鸭子，将左侧图像看作是被捕食者兔子。

Unlike animals, humans learn concepts, and these, too, can penetrate perception (Lupyan, 2017). Ambiguous figures such as Jastrow’s duck/rabbit can usually be seen under one interpretation at a time, as either a duck or a rabbit, flipping from one to the other in a moment (Lupyan, 2017). However, in a clever experiment, Jensen and Mathewson (2011) came up with verbal cues that allowed subjects to see identical pairs of ambiguous figures under both of their interpretations. Figure 5.24 shows the duck/rabbit pairs. Participants first saw these and saw them both in the usual ambiguous way, their experience flipping from one interpretation to the other. However, given the cue, “The duck is eating the rabbit,” most were able to see the right-hand image as the predatory duck and the left-hand image as the predated rabbit.

因此，一个全新的概念——食肉鸭！——能够渗透并改变这些刺激的通常感知处理（Lupyan & Clark，2015；Lupyan et al.，2020）。

Thus, a brand-new concept—carnivorous ducks!—was able to penetrate and change the usual perceptual processing of these stimuli (Lupyan & Clark, 2015; Lupyan et al., 2020).

短期记忆的特点

Characteristics of Short-Term Memory

我们有限的信息处理能力的经典证明——以及区分系统2和系统1的首要动机——来自于言语学习领域的研究——即记忆无意义的音节或单词列表。这项研究促使阿特金森和希弗林提出了两种记忆形式：短期记忆和长期记忆。

The classic demonstration of our limited capacity for information processing—and the first motive for distinguishing System 2 from System 1—comes from work in the field of verbal learning—the memorization of lists of nonsense syllables or words. This research led to Atkinson and Shiffrin’s positing of two forms of memory, the short-term store and the long-term store.

想象一下，听到并被要求记住一份以下列内容开头的列表

Imagine listening to and being asked to remember a list that starts as follows

• 长颈鹿
• Giraffe
• 茄子
• Eggplant
• 狒狒
• Baboon
• 斑马
• Zebra
• 送奶工
• Milkman
• 药剂师
• Druggist
• 香菜
• Parsley
• 萝卜
• Radish
• 打印机
• Printer
• 獭
• Otter
• 胡萝卜
• Carrot
• 牙医……
• Dentist ….

依此类推，总共 24 项。人们通常采用的学习这种清单的策略是排练，即一边听到单词一边默念。然而，在某个时候，需要记住的项目比你能承受的要多，这时就必须有所作为，产生干扰。在某些情况下，已经排练过的项目会被你设法记住的新项目所取代（前摄干扰——新知识扰乱旧知识）；在其他情况下，新项目无法进入排练集（后摄干扰——旧知识扰乱新知识）。当清单结束时，你会被要求记住尽可能多的项目，如果我们将回忆概率绘制为清单中位置的函数，我们会得到这条曲线，即序列位置效应（图 5.25）。

And so on for a total of 24 items. The strategy people usually adopt to learn this kind of list is rehearsal, saying the words silently to yourself as you hear them. However, at some point there are more items to be kept in mind than you can bear, and something’s got to give, creating interference. In some cases, items already being rehearsed get knocked out by new items you manage to latch on to (proactive interference—new learning disrupts old); in others, new items fail to enter the rehearsal set (retroactive interference—old learning disrupts new learning). When the list ends, you are asked to remember as many items as you can, and if we plot the probability of recall as a function of position in the list, we get this curve, the serial position effect (Figure 5.25).

前几项（首因效应）和后几项（近因效应）记忆良好，但中间项记忆较差。因此，在模态模型中，短期存储（STS）是信息处理的瓶颈，位于容量较大的感觉存储和长期存储之间。通过STS传递信息会导致遗忘，这要么是通过干扰（如序列位置效应所示），要么是通过衰变（Wa、Masud 和 Bays，2014），我们可以用魔术师的一种叫做时间误导的技巧来说明这一点。

The first few items (primacy effect) and the last few items (recency effect) are remembered well, but the middle items are remembered poorly. In the modal model, then, the short-term store (STS) is a bottleneck in information processing placed between the larger capacity sensory and long-term stores. Moving information through STS causes forgetting, either through interference, as demonstrated by the serial position effect, or through decay (Wa, Masud, & Bays, 2014), which we can illustrate with a magician’s technique called time misdirection.

大多数纸牌戏法都是从魔术师要求观众抽取一张牌开始的，任何一张都可以。然而，有时强制抽牌效果更好，让观众抽取一张我们选择的牌。所有强制抽牌中最简单的是交叉切牌强制抽牌¹²；它简单到你可能不相信它能起作用，但只要稍加时间误导，它就能起作用。一副牌被洗好后，魔术师会悄悄地记下底牌的牌型。参与者切牌，将牌堆的上半部分放在下半部分旁边，然后将下半部分交叉放在上半部分上，完成切牌。接下来是时间误导，魔术师描述发生了什么，指出参与者控制了牌并在任意位置切牌。然后，魔术师会通过多种可能性中的一种，揭晓被切牌选中的牌型。参与者翻开横放在牌堆另一侧的那半副牌，揭开它的身份，这与魔术师的预测相符！到现在为止，由于 STS 的容量有限以及信息衰减的趋势，参与者和观众已经忘记了“揭示”的牌只是洗牌后牌堆最底下的牌，魔术师可以轻易地知道。

Most card tricks start out with the magician asking someone to take a card, any card. However, sometimes it’s better to force a card, getting the audience member to pick a card of our choosing. The simplest of all forces is the cross-cut force¹²; it’s so simple you may not believe it works, but with a little time misdirection, it does. A deck is genuinely shuffled, and the magician quietly notes the identity of the bottom card. The participant cuts the cards, places the top half of the deck next to the lower half, and completes the cut by placing the lower half cross-wise on the upper half. Now comes the time misdirection, as the magician describes what has happened, pointing out that the participant controlled the cards and made the cut anywhere. Then, by one of many possibilities, the magician reveals the identity of the card chosen by the cut. The participant turns over the half-deck of cards resting crossways on the rest of the deck, revealing its identity, which matches the magician’s prediction! By now, because of the limited capacity and tendency for information to decay in STS, participant and audience have forgotten that the “revealed” card was just the bottom card of the deck after shuffling and could have been easily known by the magician.

从短期记忆到工作记忆：工作记忆作为长期记忆的门户和象夫

From STM to WM: Working Memory as Gateway to LTM and as Mahout

在其最初的表述中，短期记忆（或存储）被认为只是一个容量有限的通道，信息必须通过该通道进行进一步处理并存储在长期记忆中。关于其容量究竟有多有限的争论随之而来，早期公式将“神奇数字”设定为七项（Miller，1956），后来下调至四项（Cowan，2001）。然而，从一开始，人们就认为STS的意义远不止于有限的规模，因为如果运用得当，短期记忆能够容纳的神奇数字可以通过米勒所谓的“组块”进行灵活调整。在米勒看来，这是一个格式塔的理念：一些项目集合形成自然的组群，可以作为单个组块编码到记忆中，而组块在STS中只占用一个位置。

In its original formulation, short-term memory, or store, was thought of as just a limited capacity channel through which information had to pass for further processing and for storage in long-term memory. Debate ensued about how limited its capacity was, an early formulation setting the “magical number” at seven items (Miller, 1956), later revised downward to four (Cowan, 2001). However, from the beginning, there was the thought that there was more to the STS than limited size because, used properly, the magical number of items that short-term memory could hold was flexible through the deployment of what Miller called chunks. In Miller’s hands, it was a Gestalt idea: Some collections of items formed natural groups that could be encoded into memory as a single chunk taking up just one slot in STS.

想象一下，学习一个字母列表，而不是部分属于 FBIDBQ 的单词列表。它包含六个项目，接近或超过了 STS 的容量。然而，FBI 形成了一个自然组或块，由于其超出列表的含义，可以将其视为单个项目。当列表包含可分块的项目（例如 FBI）时，STS 中可以保存的项目比其他情况下更多 (Thalmann, Souza, & Oberauer, 2019)。因此，即使将相同的九个字母打乱排列，超出了短期记忆的容量，由 FBI - C - I - A - N - S - A这三个块组成的列表也可以被完全回忆起来。魔术师表演的许多最神秘的纸牌戏法都依赖于使用魔术师记住的纸牌（有时是在观众面前即兴表演！），使用分块和其他助记技巧。

Imagine learning a list of letters rather than words that goes in part F-B-I-D-B-Q. It contains six items, near or above the capacity of STS. However, F-B-I forms a natural group, or chunk, and can be thought of as a single item because of its extra-list meaning. When lists contain chunkable items, such as FBI, more items can be kept in STS than otherwise (Thalmann, Souza, & Oberauer, 2019). So, a list consisting of F-B-I-C-I-A-N-S-A, three chunks, could be completely recalled, even though the same nine letters scrambled would exceed the capacity of short-term memory. Many of the most mystifying card tricks that magicians do depend on the use of stacked decks of cards that magicians memorize (sometimes on the fly in front of the audience!), using chunking and other mnemonic techniques.

我们现在的情况

Where We Are Now

我们已经追溯了信息处理的过程，并探讨了短期记忆，它是通往长期记忆的桥梁。然而，短期记忆的概念已经发展成为一个含义更为丰富的概念，称为工作记忆（Baddeley，2016）。工作记忆不仅仅是一个信息存储库；它也是智力的关键组成部分，也是系统2的核心。接下来，我们将讨论工作记忆和长期记忆。

We have followed information processing up to short-term memory, the bridge to long-term memory. However, the concept of short-term memory has become a much richer concept called working memory (Baddeley, 2016). Working memory is much more than an information store; it’s also a key component of intelligence and the heart of System 2. Next, we move to it and LTM.

外卖

Takeaway

我们能同时进行多任务处理吗？信息处理受认知负荷的影响：我们处理的信息越多，效率和准确性就越低。例如，如果在呈现的单词列表末尾，参与者从一个随机数（例如582）开始以3为单位倒数——这是一项具有挑战性的任务——那么近因效应就会消失，因为减法任务会阻碍复习。随着数字设备的兴起，据称出现了一代精通媒体的多任务处理者，他们能够同时发短信、上课专心听讲，还能调情。13^大量研究已确凿地证明，这只是一个误区（Kirschner & De Bruyckere，2017；May & Elder，2018；Uncaphera & Wagner，2018；Lovell，2021）。这种误解之所以存在，原因之一是有些人自欺欺人地认为他们可以进行多任务处理，错误地认为他们正在进行多任务处理，而实际上他们只是在进行任务切换（Srna、Schrift 和 Zauberman，2018）。

Can we multitask? Information processing is subject to cognitive load: The more information we process the less efficiently and accurately we do it. For example, if at the end of a presented word list, participants count backward by threes from a random number (e.g., 582)—a challenging task—the recency effect disappears because the subtraction task blocks rehearsal. With the rise of digital devices, a generation of media-savvy multitaskers has allegedly arisen, adepts who can text, pay attention in class, and flirt all at the same time.¹³ A mass of research has shown conclusively that this is a myth (Kirschner & De Bruyckere, 2017; May & Elder, 2018; Uncaphera & Wagner, 2018; Lovell, 2021). One reason the myth survives is that some people fool themselves into thinking that they can multitask, falsely believing that they are multitasking when they are just task-switching (Srna, Schrift, & Zauberman, 2018).

计算和认知神经科学解释了为什么多任务处理是不可能的（Madore & Wagner, 2019; Musslick & Cohen, 2021; Rothbart & Posner, 2015），但根本问题是生物学问题。请记住，我们不可能一边做 27 × 13 的乘法一边走路。正如我们将看到的，人脑是能量消耗大户，向特别耗能的 2 号系统添加任务会导致认知崩溃。了解这些崩溃代价的绝佳资料是电视连续剧《空难调查》（2003 年以后），https://www.imdb.com/title/tt0386950/。尤其是在 1995 年左右发生的空难中，原因主要是心理因素而非机械因素，而且在几乎所有情况下，根本问题都是飞行员的信息处理能力超负荷。该剧还展示了应用认知心理学在安全工程中的应用。

There are computational and cognitive neuroscience explanations for why multitasking is impossible (Madore & Wagner, 2019; Musslick & Cohen, 2021; Rothbart & Posner, 2015), but the fundamental problem is biological. Remember that we can’t multiply 27 × 13 and walk at the same time. As we will see, the human brain is an energy hog, and adding tasks to the especially expensive System 2 results in cognitive breakdowns. An excellent source for seeing the costs of these breakdowns is the TV series Air Crash Investigation (2003+), https://www.imdb.com/title/tt0386950/. Especially in crashes occurring after about 1995, the causes are primarily psychological rather than mechanical, and, in almost every case, the root problem is overloading pilots’ information processing limits. The show also showcases jobs for applied cognitive psychology in safety engineering.

有时，让认知负荷发挥作用是可能的。在康复初期，我服用了一种可能产生幻觉副作用的药物。我开始“听到”妻子呼救，但这根本不可能。所以，当这种情况发生时，我会倒数三，幻觉就会消失。这就引出了——

Sometimes it’s possible to make cognitive load work for you. Early in rehab I was put on a medication that had a potential side effect of inducing hallucinations. I started “hearing” my wife calling out for help, which was impossible. So, when it happened, I’d count backward by threes, and the hallucination would vanish. This takes us to—

阅读材料

Readings

唐纳德·霍夫曼从预测加工的角度撰写了《反现实案例：进化为何将真相从我们眼前隐藏》（纽约：WW·诺顿出版社，2019年）。他反驳了福多尔所认可的“心智是为了寻找真理”的哲学论点，认为了解什么是适应性更为重要。

Donald Hoffmann makes The Case against Reality: Why Evolution Hid the Truth from Our Eyes (New York: W. W. Norton, 2019) from the predictive processing perspective. He argues against the philosophical thesis (endorsed by Fodor) that the mind is for finding Truth, arguing that knowing what’s adaptive is more important.

理查德·格雷戈里是一位研究感知的领军人物，兴趣广泛，兼收并蓄。他写了两本引人入胜的书，分别是《奇异感知》和《更奇异感知》。 （纽约：劳特利奇出版社，1989年，1994年），运用感知的怪癖来阐释其心理学。他编辑了一部作品《巧妙的眼睛》（牛津：牛津大学出版社，1995年），书中介绍了顶尖心理学家如何将科学应用于艺术，包括维米尔对暗箱的运用。

Richard Gregory was a leading perception researcher of broad and eclectic interests. He wrote two fascinating books, Odd Perceptions and Even Odder Perceptions (New York: Routledge, 1989, 1994), using quirks of perception to illustrate its psychology. He edited a collection, The Artful Eye (Oxford: Oxford University Press, 1995), in which leading psychologists applied science to art, including Vermeer’s use of the camera obscura.

蒂莫西·威尔逊 (Timothy Wilson) 的《陌生人》（2002 年）总结了关于我们对大象活动过程缺乏认识的研究，但他将这些研究结果用于实际目的，例如如何更快地入睡。

Timothy Wilson’s Strangers to Ourselves (2002) contains summaries of research on our poor awareness of the elephant’s processes, but he uses the findings for practical ends, such as how to go to sleep more quickly.

电影

Movies

http://www.tasteofcinema.com/2015/10-great-movies-that-question-the-nature-of-human-perception/上列出了一些相关的电影。类似的网站还有https://www.filminquiry.com/6-movies-screw-sense-reality/，但它指的是罗森汉一项现已被否定的关于精神错乱的研究。

At http://www.tasteofcinema.com/2015/10-great-movies-that-question-the-nature-of-human-perception/ there’s a list of relevant films. Similar is https://www.filminquiry.com/6-movies-screw-sense-reality/, but it refers to a now-discredited study on insanity by Rosenhan.

认知心理学家研究电影体验（例如，电影中对phi现象的运用）。两篇总结性文章如下：ES Tan，“电影心理学”，Palgrave Communications 4 (2018): 第82篇。DOI: 10.1057/s41599-018-0111-y | www.nature.com/palcomms；以及EB Goldstein，“电影感知”，载于《感知百科全书》第一卷（洛杉矶：SAGE出版社，2010年），第459-461页。DOI: http://dx.doi.org/10.4135/9781412972000.n146。

Cognitive psychologists study the movie experience (e.g., its use of the phi phenomenon). Two summary articles are: E. S. Tan, “A Psychology of the Film,” Palgrave Communications 4 (2018): Article 82. DOI: 10.1057/s41599-018-0111-y | www.nature.com/palcomms; and E. B. Goldstein, “Film (Cinema) Perception,” in Encyclopedia of Perception, Vol. 1 (Los Angeles: SAGE, 2010), pp. 459–461. DOI: http://dx.doi.org/10.4135/9781412972000.n146.

最后，网站https://www.k-state.edu/psych/vcl/applied-research/film.html由一个广泛研究电影认知心理学的实验室维护，包括感知、注意力和记忆。

Finally, the website https://www.k-state.edu/psych/vcl/applied-research/film.html is maintained by a lab investigating the cognitive psychology of film broadly, including perception, attention, and memory.

魔法

Magic

在众多可能性中，本章给出了两个例子。

Out of many, many possibilities, here are two examples for this chapter.

图像的持久性被用来创造魔术师所谓的“视觉保留” ——手中的小物体（例如硬币或海绵球）消失。魔术师一只手拿着物体，假装把它放在另一只手上。由于视觉保留，观众会暂时“看到”物体在接球的手中，但当物体不在时，他们会感到惊讶。演示和教程可在https://www.youtube.com/watch?v=JdxWuLOHaJ8上观看。想要观看我见过的最好且价格实惠的教程，请访问https://www.youtube.com/watch?v=_8sPAu2o99A。

The persistence of the icon is exploited to create what magicians call retention of vision vanishes of small objects in the hand, such as coins or sponge balls. The magician holds the object in one hand and pretends to place it in the other. Because of retention of vision, the audience “sees” the object momentarily in the receiving hand and is surprised when it’s not there. Demonstrations and a tutorial are available at https://www.youtube.com/watch?v=JdxWuLOHaJ8. For the best I’ve ever seen and an inexpensive tutorial, visit https://www.youtube.com/watch?v=_8sPAu2o99A.

魔术师们毫不留情地利用了人们对于变化视盲的现象。公主牌戏法就是一个很好的例子。

Magicians mercilessly exploit change blindness. A case in point is the Princess Card Trick.

通常情况下，公主牌戏法是用一副五张牌的牌组为一位观众表演的，但我使用的是为团体表演而开发的改良版本。你将学习这个戏法的认知秘诀，但不会学习牌组操作的秘诀。在图5.26中，你会看到五张扑克牌的展示。快速地看一眼（不超过五秒钟左右），闭上眼睛，想出一张要从牌组中剔除的牌。值得注意的是，今天，也就是2019年7月29日，我已经预测了你会选哪张牌，并且安排了四张牌的展示，但少了你的牌。请翻到图5.27，看看我是否猜对了。你的牌在那儿吗？

Normally, the Princess Card Trick is performed with a packet of five cards for a single spectator, but I’m using a modified version developed for groups. You will learn the cognitive secret of the trick but not the card manipulation secret. In Figure 5.26 you will find a display of five playing cards. Look at it briefly (no more than five seconds or so), close your eyes, and think of one card to eliminate from the packet. Remarkably, today, July 29, 2019, I have predicted which card you will choose, and have arranged a display of four cards that’s missing yours. Turn to Figure 5.27 and see if I’m right. Is your card there?

我怎么会提前这么久知道你的选择？答案是视盲。请看图 5.28。最上面一排 5 张牌是最初展示的牌，最下面一排是我从中“移除”了一张牌的牌。但是，这些牌都不在第一个展示牌中，因此无论你从第一个展示牌中选择什么牌，它都不在四张牌的牌组中。但是，由于扑克牌看起来相当相似，你没有注意到从第一个五张牌展示牌到最后一个五张牌展示牌的变化，所以以为我以某种方式猜出了你的牌。你对从第一次展示牌到第二次展示牌阵的变化视盲。

How did I know your choice so far in advance? The answer is change blindness. Look at Figure 5.28. The top row of 5 cards was the initial display, and the bottom row is the display from which I had “removed” one card. However, none of these cards was in the first display, therefore no matter what card you chose from the first display, it was not in the four-card set. However, since playing cards look rather similar, you did not notice the change from the first five-card display to the last, and so thought I had somehow divined your card. You were blind to the change in the card array from the first presentation to the second.

下一章记忆演示。请阅读以下段落一两遍，但不要重复。下一章将要求你尽力回忆这段文字及其标题。

Memory Demonstration for Next Chapter. Please read the following passage once or twice, but do not return to it. In the next chapter you will be asked to recall the passage and its title as best you can.

记忆是一种记录。不：记忆是对过去的重建

Memory Is a Recording. No: Memory Is a Re/construction of the Past

奥古斯丁将记忆比作洞穴的比喻，在20世纪不可避免地被“记忆是记录”的理念所取代（Draisma，2000）。不幸的是，记录的比喻强化了人们对记忆的直觉性信任。尽管目前人们对深度伪造视频和Photoshop™的担忧引发了人们对记忆的质疑，但记忆就像一段可以倒带重放的录音的想法依然存在。录音机的比喻保留了奥古斯丁的空间特质，暗示记忆是一个被动的仓库（Roediger，1980），而正如我们将看到的，它更像是橡皮泥。

Augustine’s metaphor for memory as caves and caverns was inevitably replaced in the 20th century by the idea of memory as a recording (Draisma, 2000). Unfortunately, the recording metaphor reinforces intuitive faith in memory. Although at present skepticism about memory is provoked by worries about deep fake videos and Photoshop™, the idea that memory is like a recording that can be rewound and replayed persists. The recorder metaphor retains the spatial quality of Augustine’s, suggesting that memory is a passive storehouse (Roediger, 1980), when as we’ll see, it’s more like silly putty.

我们将了解到，记忆是对过去的重构。即使这种表述也过于薄弱，因为正如我们刚刚了解到的，我们对世界的持续体验是一种即时构建，将我们预测的世界模型构建与来自感官的纠错信号融合在一起。因此，即使记忆是完美的记录，它也是对所发生事件的解读的完美记录，而不是对已发生事件的记录。此外，每一次记忆行为都会使用相同的构建过程来创造新的体验，但却没有纠错信号的可能性。普鲁斯特曾说过：“对过去事物的回忆不一定是对事物本来面目的回忆。”

Memory, we will learn, is a reconstruction of the past. Even this formulation is too weak because, as we just learned, our ongoing experience of the world is an on-the-fly construction merging our predictive world model-building with error correction signals from the senses. Therefore, even if memory were a perfect recording, it would be a perfect recording of an interpretation of what happened, not a recording of what did happen. In addition, every act of remembering creates a new experience using the same constructive processes but without the possibility of error correction signals. Proust remarked, “Remembrance of things past is not necessarily remembrance of things as they were.”

一切都在那里。不：存在遗忘、干扰和记忆的易变性

It’s All in There. No: There Is Forgetting, Interference, and Mutability of Memories

关于记忆最有害的迷思之一是：如果记忆是一种记录，那么我们所经历的一切都会被储存在某个地方。弗洛伊德相信这一点，这导致他犯下了灾难性的诱惑错误，并由此创立了精神分析学（Leahey，2018）。一些现代治疗师坚持相信这一点，导致了道德恐慌和对虐童的诬告。催眠师也相信这一点，这导致他们通过暗示性探寻谋杀和外星人绑架的记忆。

One of the most pernicious myths about memory is that if memory is a recording, then everything we’ve ever experienced is stored away somewhere. Freud believed it, leading him to make the disastrous seduction mistake that created psychoanalysis (Leahey, 2018). Some modern therapists persisted in believing it, leading to moral panics and false accusations of child abuse. Hypnotists believe it, leading to suggestive probing for memories of murders and alien abductions.

然而，正如我们在感觉记忆和短期记忆（STM）研究中发现的那样，并非所有记忆都包含在内。有些体验由于回溯性干扰而无法编码，而另一些体验则由于前摄性干扰而丢失。未被注意到的体验部分，例如意外遇到的大猩猩，根本就不会被编码。

However, as we have already found from research on sensory and short-term memory (STM), it’s not all in there. Some experiences are blocked from encoding because of retroactive interference, and others are lost because of proactive interference. Unattended parts of experience, such as unexpected gorillas, never get encoded at all.

记忆是储存的经验。不：记忆是重新绑定的经验

Memories Are Stored Experiences. No: Memories Are Re-bound Experiences

无论是奥古斯丁的洞穴，还是如今的计算机文件，存储的隐喻都意味着每段记忆都存储在心理或神经空间的某个特定区域。例如，大脑中某个文件夹里保存着你昨晚晚餐的记忆，或者你第一只宠物的神经痕迹。正如奥古斯丁所说：“它的形象仍然保留在大脑中，人们不断寻找它，直到它重回视野。” 然而，事实并非如此。感知的绑定问题也适用于记忆。过去经验原本绑定的感官表征必须重新绑定到过去的某个事件中。

The storage metaphor, whether Augustine’s caverns or today’s computer files, implies that each memory is stored in a single part of psychological or neural space. So, for example, there’s a folder holding your memory of last night’s dinner or a neural trace of your first pet somewhere in the brain. As Augustine said, “its image is still retained within, and it is sought until it be restored to sight.” However, this is not true. The binding problem of perception applies to memory, too. The originally bound sensory representations of past experience must be re-bound into a single event from the past.

记忆是一回事。不：“记忆”可以分解成不同的类型

Memory Is One Thing. No: “Memory” Can Be Decomposed into Types

我们已经根据记忆存储于信息处理中的位置将其分为不同的类型。长期记忆（LTM）本身也可以进行类似的划分。

We have already divided memory stores into types based on their position in information processing. Long-term memory (LTM) itself can likewise be divided.

语义记忆与情景记忆（自传体记忆）

Semantic vs. Episodic (Autobiographical) Memory

在艾宾浩斯（Ebbinghaus）创立的言语学习范式（特别是关于词汇表学习）的背景下，图尔文（Tulving，1972）提出，长期记忆（LTM）存在不同形式，这些形式由其心理功能和结构定义。图尔文认为，学习词汇表涉及两个记忆系统。首先，实验参与者会创造一段关于他们生活中新事件的记忆；其次，他们会运用一种已经存在的记忆形式，即对词汇表单词含义的记忆。因此，他将第一种记忆称为情景记忆，将第二种记忆称为语义记忆。

In the context of the verbal learning paradigm invented by Ebbinghaus, and specifically about word-list learning, Tulving (1972) proposed that there were different forms of LTM defined by their psychological functions and structure. Tulving argued that learning a list of words involved two memory systems. First, being participants in an experiment created a memory of a new episode in their lives; second, they were deploying an already existing form of memory, memory for the meaning of the words in the list. Hence, he dubbed the first kind of memory episodic memory, and the second, semantic memory.

情景记忆的组织方式是时间性的。当我们搜索它时，我们会沿着我们生活的时间线进行搜索，找出所需记忆在一系列其他记忆中发生的时间。所以，如果我问你上周二下午1点在哪里，你可能会回忆起你的课程安排和社交互动，最终想起上完心理学课后，你和朋友在一家露天咖啡馆共进午餐。如今，“自传体记忆”这个术语已经在很大程度上取代了图尔文最初提出的“情景记忆”。

The organization of episodic memory is temporal. When we search it, we search on a timeline of our lives, working out when a requested memory occurred in a sequence of other memories. So, if I ask you where you were last Tuesday at 1 PM you might recall your class schedule and social interactions to finally retrieve that after psych class you were having lunch with a friend at an outdoor café. Today, the term autobiographical memory has largely replaced Tulving’s original episodic memory.

语义记忆的组织结构由词义构成。例如，当回忆“长颈鹿”、“茄子”、“狒狒”、“送奶工”这样的列表时，参与者倾向于按语义类别对条目进行分组，例如长颈鹿和狒狒，送奶工和牙医，等等。我们很少在记忆中为单词添加可时间定位的标签（你可能不知道自己是什么时候学到长颈鹿的），但在某些特殊情况下，你可能会这样做（比如你昨天去看了牙医）。历史可以改变记忆：你曾经遇到过送奶工吗？（我家遇到过。）打印机是工作还是设备？

The organization of semantic memory is structured by the meaning of words. For example, when recalling a list such as the giraffe, eggplant, baboon, milkman list, participants tend to group the items by semantic category, giraffe with baboon, milkman with dentist, and so on. We rarely have temporally locatable tags on memory for words (you probably don’t know when you learned what a giraffe was), though in some idiosyncratic cases you may (you saw the dentist yesterday). History can alter memorability: Have you ever had a milkman? (My family did.) Was printer a job or a device?

陈述性记忆与程序性记忆

Declarative vs. Procedural Memory

20世纪40年代，英国哲学家吉尔伯特·赖尔（Gilbert Ryle，1946年，1949年）区分了两种类型的知识：知识如何（knowing-how）和知识如何（knowing-that）。前者指的是学习到的活动，例如骑自行车、系鞋带或做煎蛋卷。知识如何指的是我们已知的事实，例如“巴黎是法国的首都”或“煎蛋卷是用鸡蛋做的”。重要的是，这两个类别之间存在重叠，但不完全重叠。NBA球星可以投进三分球，但他们可能无法描述其中涉及的物理原理——他们拥有知识如何，但没有知识如何——而物理学家通常恰恰相反，拥有知识如何，却没有知识如何。在某些情况下，知识如何和知识如何可能部分重叠。我可以教你如何做煎蛋卷，直到把它抛成最终的形状，这需要实际的运动技能学习；可以演示，但你必须练习^。1

In the 1940s, British philosopher Gilbert Ryle (1946, 1949) drew a distinction between two types of knowing, knowing-how and knowing-that. The first referred to learned activities such as riding a bicycle or tying one’s shoes or making omelets. Knowing that referred to things we know as facts, such as “Paris is the capital of France,” or “Omelets are made with eggs.” Importantly, the categories overlap, but incompletely. NBA stars can sink three-point shots, but they probably can’t describe the physics involved—they have know-how without knowing-that—while physicists are usually the reverse, having know-that without the know-how. In some cases, it’s possible for knowing-how and knowing-that to overlap partially. I can teach you how to make an omelet up to the point of tossing it into its final form, which takes actual motor-skill learning; it can be shown, but you must practice.¹

Squire & Zola-Morgan (1988) 借鉴临床和实验神经科学，对程序性学习（或记忆）和陈述性学习（或记忆）进行了类似的区分。程序性学习是指你能做的，而陈述性学习是指你能谈论的。从某种意义上说，这种区分是赫尔和托尔曼学习理论的核心区别。赫尔强调做/程序，而托尔曼强调陈述/知识。由于缺乏赖尔或斯奎尔的区分，他们认为这场战斗是“非此即彼”，而不是实际上的“两者兼而有之”。

Drawing on clinical and experimental neuroscience, Squire & Zola-Morgan (1988) made a similar distinction, between procedural learning (or memory) and declarative learning (or memory). Procedural learning is what you can do, declarative learning is what you can talk about. In a sense, this distinction was the core difference between Hull’s and Tolman’s theories of learning. Hull emphasized doing/procedure, while Tolman emphasized declaring/knowledge. Lacking either Ryle’s or Squire’s distinctions, they saw their battle as either-or rather than, as it really was, both-and.

外显记忆与内隐记忆

Explicit vs. Implicit Memory

LTM 记忆类型的最后一个主要分类是外显记忆和内隐记忆(Graf & Schacter, 1985)。目前，最好将其作为对各种记忆实验的描述性、方法论性区分 (Schacter, 1987)。外显记忆是人们有意识地提取的记忆，而内隐记忆则是那些在自身未被意识到的情况下影响行为的记忆。

The final major taxonomy of LTM memory types is between explicit and implicit memory (Graf & Schacter, 1985). For now, it’s best to present this as a descriptive, methodological distinction about kinds of memory experiments (Schacter, 1987). Explicit memories are those that people consciously retrieve, while implicit memories are memories that influence behavior without themselves coming to awareness.

内隐记忆的例子来自认知启动研究。让我们回到长颈鹿、茄子、狒狒的列表来进行说明。在使用这些物品的启动实验中，我们向参与者逐个显示计算机屏幕上的单词。我们要求他们在看到每个单词时尽快说出它，并测量每个单词开始被说出所需的时间：它的延迟或反应时间。我们可以操纵单词的前后呈现之间的关系。例如，假设第三个单词是“狒狒”，如上所示。最初呈现的单词是启动，而后面的单词是目标。我们感兴趣的是看看启动是否促进（引发）对目标的反应。

Examples of implicit memory come from studies of cognitive priming. Let’s return to the giraffe, eggplant, baboon list to illustrate. In a priming experiment with these items, we present participants with the words one at a time on a computer screen. We ask them to say each word as quickly as possible when they see it, and we measure how long it takes for each word to begin being said: its latency or reaction time. We can manipulate the relationship between earlier and later presentations of the words. Suppose, for example, the third word is “baboon,” as above. The initially presented word is the prime and the later word is the target. We are interested in seeing if the prime facilitates (primes) response to the target.

现在想象第六个单词是这些中的一个：狒狒；狒狒；黑猩猩；狮子；打印机。我们通常会发现，与初始呈现的反应时间相比，身份启动狒狒（相同字体的相同单词）的 RT 最短；狒狒（不同字体的相同单词）稍长一些；语义启动黑猩猩（相关物种）和狮子（另一种丛林动物）稍长一些；与对照组相比没有变化，对照组的第一个单词是语义上不相关的目标打印机。因为呈现时间很短，所以参与者通常不会注意到启动和目标之间的关系，从而使启动成为内隐记忆的问题。

Now imagine that the sixth word is one of these: baboon; baboon; chimpanzee; lion; printer. We will typically find that compared to the reaction time for the initial presentation, the RT will be shortest for the identity prime baboon (same word in same typeface); a bit longer for baboon (same word in different typeface); a bit longer for the semantic primes chimpanzee (related species) and then lion (another jungle animal); and unchanged compared to controls where it was the first word, the semantically unrelated target printer. Because presentation times are brief, participants usually don’t notice the relationships between primes and targets, making priming a matter of implicit memory.

我们可以通过改变任务来强化启动效应的内隐、无意识影响。参与者首先像在短期记忆实验中一样学习一串单词，并接触潜在的启动效应。为了验证这一点，他们随后进行一个单词片段补全任务，填补单词片段中缺失的空格，例如 b__o_n。一组（外显记忆组）被告知部分单词片段来自之前的记忆列表；另一组（内隐记忆组）则没有被告知。此外，两组都看到了填充词，也就是说，这些单词之前没有出现过，以减少内隐组注意到单词重复的机会。然而，两组都对重复的单词（即启动词）表现出了更强的理解能力。最有趣的是，患有失忆症的受试者在完成单词片段拼图时会完全忘记参与初始列表学习任务，但他们对目标词仍然会表现出正常或接近正常的启动（Graf & Schacter, 1985）。他们失去了形成外显记忆的能力，但他们仍然会受到他们没有意识到的刺激的正常启动。

We can strengthen the case for implicit, unconscious influence of primes by changing the task. Participants first learn a list of words as in a short-term memory experiment, exposing them to potential primes. To test this, they later do a word fragment completion task, filling out missing spaces in word fragments such as b__o_n. One group, the explicit memory group, is told that some of the word fragments came from the earlier memory list; the other, the implicit memory group, is not told. In addition, both groups see filler words, that is, words not presented earlier, to reduce the chance the implicit group notices the word repetitions. However, both groups show enhanced solution for the repeated, i.e., primed words. Most interestingly, subjects with amnesic syndrome, who by the time of doing the word fragment puzzle will have completely forgotten participating in the initial list learning task, will nevertheless display normal to near normal priming for the target words (Graf & Schacter, 1985). Their ability to form explicit memories is lost, but they continue to be normally primed by stimuli of which they are not aware.

记忆的首次研究

The First Study of Memory

艾宾浩斯于 1885 年出版了《论记忆》 （Über das Gedachtnis）。与当时大多数心理学家一样，艾宾浩斯采用了联想主义的思维概念，因此他认为联想的形成是学习和记忆的基本过程。他将联想称为连接思想的“心理线索”。他想研究这些线索以及它们如何独立于内容来连接和解开思想。因此，他拒绝研究单词、诗歌或段落等有意义的材料，因为他认为人们学习、记忆和遗忘它们的方式会混淆理解的认知过程和联想学习。实际上，艾宾浩斯试图解开记忆的戈耳迪之结；像亚历山大一样，他成功了，但没有在关节处雕刻出自然的形状。

Ebbinghaus published On Memory (Über das Gedachtnis) in 1885. Like most psychologists of the time, Ebbinghaus adopted the associationist conception of the mind, and thus he saw association formation as the basic process of learning and memory. He called associations “mental threads” that connect ideas. He wanted to study the threads and how they bind and unbind ideas independently of their content. He therefore rejected studying meaningful material such as words, poems, or paragraphs, because how people learn, remember, and forget them would confuse, as he thought, cognitive processes of comprehension with association learning. Ebbinghaus, in effect, sought to cleave the Gordian knot of memory; like Alexander, he succeeded at the cost of not carving nature at the joints.

为了检验他的线索，艾宾浩斯发明了无意义音节，即在两个辅音之间放置一个元音：TOB、SAB、GEN等等。这些音节很容易构造，并能产生大量可学习的材料。由于它们毫无意义，需要纯粹的记忆，这就排除了兴趣、与已学内容的相关性以及对意义的理解等因素。为了进行实验控制，艾宾浩斯构建了音节列表，并以恒定的速度在一天中的固定时间呈现。整个过程很容易量化：艾宾浩斯可以测量他（他的第一个参与者）在改变自变量时学习或重新学习音节列表所花费的时间，希望像巴甫洛夫一样，发现联想学习的科学规律。

To examine his threads, Ebbinghaus invented the nonsense syllable, constructed by placing a vowel between two consonants: TOB, SAB, GEN, and so on. They were easy to make up and produced a large amount of learnable material. Being meaningless, they required sheer memorization, ruling out such factors as interest and relevance to what has already been learned as well as comprehension of meaning. To impose experimental control, Ebbinghaus constructed lists of syllables to be presented at a constant rate of speed and at consistent times of the day. The whole procedure was easily quantifiable: Ebbinghaus could measure how long it took him (his first participant) to learn or relearn lists of syllables as he changed independent variables, hoping, like Pavlov, to discover scientific laws of associative learning.

就启发价值而言，艾宾浩斯的方法取得了巨大的成功，因此它在人类学习中长期占据主导地位。我们只讨论其中一项发现，即艾宾浩斯遗忘曲线。关于记忆的一个最明显的事实是，我们学的东西越久远，就越难记住。根据个人经验，我们只能这么说。但艾宾浩斯的实验使得更精确地提出问题并获得定量答案成为可能。在一系列实验中，艾宾浩斯记住了八组包含 13 个无意义音节的音节表；之后，在不同的时间之后，他再次记住了它们。因此，艾宾浩斯可以根据重新学习的难度准确地问出：在给定时间内，一组无意义音节中有多少被忘记了？

As far as heuristic value goes, Ebbinghaus’ method was a smashing success, hence its dominance of human learning for so long. We will look at just one finding, the Ebbinghaus curve of forgetting. One of the most obvious facts about memory is that the longer ago we learned something, the harder it is to remember. Based on personal experience, we can say no more than that. But Ebbinghaus’s experiments made it possible to frame the question more precisely and receive a quantitative answer. In a series of experiments, Ebbinghaus learned by heart eight lists of 13 nonsense syllables each; later, after differing amounts of time, he learned them by heart again. Consequently, Ebbinghaus could ask precisely, in terms of difficulty of relearning: How much of a list of nonsense syllables is forgotten at a given time?

图6.1展示了艾宾浩斯实验的结果；它显示了遗忘材料的百分比随初始学习后小时数的变化。大多数遗忘发生得很快。学习后仅一小时，超过55%的无意义音节就被遗忘了，而超过8.8小时的额外遗忘在31天后也只有14%。艾宾浩斯准确地展示了遗忘发生的速度。艾宾浩斯的遗忘曲线简直是老师的噩梦：课后几个小时，学生们就忘记了大部分所教的内容！

Figure 6.1 illustrates the results of Ebbinghaus’ experiments; it shows the percent of material forgotten as a function of hours after initial learning. Most forgetting takes place quickly. Only an hour after learning, more than 55 percent of the nonsense syllables have been forgotten, while the additional loss past 8.8 hours is only 14 percent after 31 days. Ebbinghaus showed exactly how quickly forgetting takes place. Ebbinghaus’ curve of forgetting is a teacher’s nightmare: Just hours after class, students will have forgotten most of what was taught!

另一方面，并​​非所有记忆都会以如此惊人的速度衰退。例如，在一项巧妙的实验中，Bahrick、Bahrick 和 Wittlinger（1975）利用高中毕业纪念册，测量参与者从毕业两周到54年后对同学姓名和面孔的记忆。对于姓名，他们采用了自由回忆和再认记忆法。在自由回忆中，参与者被要求记住尽可能多的同学姓名。为了测试再认记忆，他们要求同学姓名以名单与非同学的对照姓名混杂在一起。对于图片，识别记忆的测量方法是将同学的年鉴照片与同期其他高中年鉴中的对照照片混杂在一起；参与者需表明是否记得每张照片中的人。在线索回忆程序中，自由回忆被部分模仿，该程序会给出图片的线索，并询问参与者是否记得这些学生。Bahrick等人发现，即使在最长的时间内，对姓名和面孔的识别记忆也几乎没有下降。姓名的自由回忆和图片的线索回忆确实会随着时间的推移而下降，并且与艾宾浩斯曲线的形状一致，尽管速度要慢得多。对无意义音节的记忆仅需几个小时，而对姓名和面孔的记忆则需要几十年。一般而言，视觉记忆（尤其是通过识别评估的视觉记忆）比言语记忆要强得多，对朋友等重要社会信息的记忆也是如此。

On the other hand, not all memories decay at such an alarming rate. For example, in a clever experiment, Bahrick, Bahrick, and Wittlinger (1975) used high school yearbooks to measure participants’ memory of names and faces of fellow students from two weeks to 54 years after graduation. For names, they used free recall and recognition memory methods. In free recall, participants are asked to remember as many fellow students’ names as possible. To test recognition memory, fellow students’ names were shown in lists mixed up with control names of non-fellow students. For pictures, recognition memory was measured by yearbook pictures of fellow students mixed up with control pictures from yearbooks of other high schools from the same period; participants indicated whether they remembered the person in each picture. Free recall was partially imitated in a cued recall procedure, in which cues to the pictures were given, and participants were asked if they recalled the students. Bahrick et al. found that recognition memory for both names and faces scarcely declined at all even over the longest periods. Free recall of names and cued recall of pictures did decline over time and in accord with the shape of Ebbinghaus’ curve, though much more slowly. What took mere hours for nonsense syllables took decades for names and faces. In general, visual memory, especially as assessed by recognition, is much more robust than verbal memory, as is memory for socially important information such as friends.

Mahout的工作记忆

Working Memory as Mahout

一些批评者拒绝信息处理模态模型的原因之一是，区分短期记忆和长期记忆有时会遇到困难（Ma, Husain, & Bays, 2014; Shevlin, 2020；另见Norris, 2017）。例如，定义短期记忆（STM）的现象之一是序列位置效应，但它也出现在长期记忆（LTM）中。一些研究要求美国人自由回忆他们的总统，并发现了强烈的序列位置效应（DeSoto & Roediger, 2019）。只有林肯的总统任期在中间位置突出（图6.2）。

One of the reasons why some critics reject the modal model of information processing is the occasional difficulty of distinguishing short-term and long-term memory (Ma, Husain, & Bays, 2014; Shevlin, 2020; but see Norris, 2017). For example, one of the phenomena that defines STM is the serial position effect, but it appears in LTM, too. Several studies have asked Americans to free-recall their presidents and found a strong serial position effect (DeSoto & Roediger, 2019). Only Lincoln’s presidency stands out in the middle (Figure 6.2).

此外，所有形式的记忆都可以进行策略管理，从而模糊了短期记忆 (STM) 和长期记忆 (LTM) 之间的界限。一个重要的例子是前瞻性记忆(Brandimonte, Einstein, & McDaniel, 1996; Kliegel, McDaniel, & Einstein, 2008; MacBride & Workman, 2017; McDaniel & Einstein, 2007)。当我们思考记忆时，我们会想到回顾性记忆，就像奥古斯丁和普鲁斯特所探讨的那种。然而，我们也以不同的方式使用长期记忆，就像在购物清单中列出待办事项一样。回顾性记忆和前瞻性记忆并非旨在作为两种不同的记忆形式，而是作为使用现有长期记忆的不同方式，而使用前瞻性记忆需要谨慎的策略管理。我们必须列出清单，记住要参考它，并监控我们的行为，注意在正确的时间按正确的顺序做事，并且不要忘记任何一件事。²它涉及元记忆——思考记忆——就像我们离开家时问自己“我关掉烤箱了吗？”，或者反思我们个人的记忆故障，“我不擅长将名字与面孔匹配”，或衰老的影响，“这些天我经常处于舌尖状态。”你可以想象，有效地运用前瞻性记忆在现实生活中非常重要（Dismukes，2012；Walter＆Meier，2014）。

Moreover, all forms of memory can be strategically managed, blurring the distinction between STM and LTM. An important example is prospective memory (Brandimonte, Einstein, & McDaniel, 1996; Kliegel, McDaniel, & Einstein, 2008; MacBride & Workman, 2017; McDaniel & Einstein, 2007). When we think of memory, we think of retrospective memory, the kind probed by Augustine and Proust. However, we also use LTM in a different way, as providing a list of things to do as in a shopping list. Retrospective and prospective memory are not intended as separate forms of memory, but as different ways of using existing LTM, and using prospective memory requires careful strategic management. We must form the list, remember to consult it, and monitor our behavior, being careful to do things in the right order at the right time and not to forget any of them.² It involves metamemory—thinking about memory—as when we ask ourselves, “Did I turn off the oven?” when leaving the house, or reflecting on our personal memory glitches, “I’m bad at matching names to faces,” or effects of aging, “I get in tip-of-the-tongue states more often these days.” As you can imagine, effective prospective memory use is important in real life (Dismukes, 2012; Walter & Meier, 2014).

诸如此类的考量表明，STM 不仅仅是一个容量有限的盒子，用于短暂地存储输入的想法，它还管理着 Atkinson 和 Shiffrin 模态模型中提到的控制过程，从而催生了工作记忆(WM) 的概念。迄今为止，最具影响力的工作记忆理论是艾伦·巴德利 (Alan Baddeley) 的理论。巴德利提出的该模型是一种非正式的模型，旨在指导研究，但受到了更明确的计算方法的挑战，例如 Oberauer (2009; 2019; 详见 Shevlin, 2020)。为了便于我们探讨，我们将继续使用巴德利的模型。

Considerations such as these indicated that STM was more than simply a limited capacity box in which incoming ideas could be briefly placed, but also managed the control processes mentioned in the Atkinson and Shiffrin modal model, creating the concept of working memory (WM). By far the most influential theory of working memory is Alan Baddeley’s. Baddeley proposed the model as an informal one meant to guide research and has been challenged by more explicit computational approaches, such as Oberauer (2009; 2019; for review, Shevlin, 2020). For our purposes, we will stick with Baddeley.

Baddeley (2012) 提出的模型由几个部分组成（图 6.3）。Baddeley 从智力心理测量学中引入了流体系统和晶体系统之间的区别。智力的一个公认分类是流体智力和晶体智力。晶体智力大致就是你所知道的东西。它有两个重要组成部分：情景记忆（我们对自己过去的了解）和语义记忆（我们对思想含义的认识）。后者对智力的重要性超乎你的想象。例如，智商测试的词汇分项与整体智商的相关性约为 0.80。

Baddeley (2012) presents his model as composed of several components (Figure 6.3). Baddeley imported the distinction between fluid and crystallized systems from the psychometrics of intelligence. One widely acknowledged division of intelligence is between fluid intelligence and crystallized intelligence. Crystallized intelligence is roughly what you know. Its two important components are episodic memory, what we know of our own past, and semantic memory, our knowledge of the meaning of ideas. The latter is more important for intelligence than you might think. For example, the vocabulary subtest of IQ tests correlates about .80 with overall IQ.

流体智力与解决问题有关。它包含三个缓冲区，提供感官输入和长时记忆（LTM，晶体智力）之间的接口。视觉空间速写板将视觉记忆和视觉输入汇集到一个可以整合和操控它们的地方，就像你在视觉记忆中搜索特定信息一样。语音回路对听觉记忆和输入也起着同样的作用，就像你复述“长颈鹿……茄子……狒狒……”这样的列表一样。情景缓冲区保存绑定的情景——即在我们通常的多模态体验中，感官已经结合在一起的输入和记忆——用于短暂的存储和操控，因此是当前输入与新情景记忆形成之间的关键接口（Baddeley、Allen & Hitch，2011；Popov、Marevic、Rummel & Reder，2019）。中央执行器负责监督这些容量有限的缓冲区，它管理整个系统的信息处理。

Fluid intelligence is about problem solving. It contains three buffers that provide interfaces between sensory input and LTM (crystallized intelligence). The visuo-spatial sketchpad brings together visual memories and visual inputs into a place where they can be integrated and manipulated, as when you search a visual memory for a specific bit of information. The phonological loop does the same for auditory memories and inputs, as when you rehearse a list such as Giraffe … Eggplant … Baboon…. The episodic buffer holds bound episodes—that is, inputs and memories in which the senses have been united in our usual multi-modal experiences—for brief storage and manipulation and is therefore the key interface between current input and the formation of new episodic memories (Baddeley, Allen, & Hitch, 2011; Popov, Marevic, Rummel, & Reder, 2019). Supervising these limited capacity buffers is the Central Executive, which manages information processing in the system as a whole.

以下是工作记忆在双任务跨度任务中的运用示例（Daneman & Carpenter, 1980），因为两个过程必须在一段时间内同时进行（Hunt, 2011）。参与者会看到一些句子，并需要重复每个句子的最后一个单词；完成这些句子后，参与者需要回忆每个句子的最后一个单词。

Here is an example of working memory in use (Daneman & Carpenter, 1980) in what is called a dual-task span task because two processes must be simultaneously performed over a span of time (Hunt, 2011). The participant is shown a number of sentences and has to repeat the last word of each sentence; when the sentences are done, the participant is asked to recall the pronounced last word of each sentence.

呈现的句子：

Sentences presented:

• 他顶着凛冽的寒风沿着密歇根大道走去。
• He walked down Michigan Avenue in the face of the bitter wind.
• 该餐厅因其富有创意的鱼类菜肴而闻名。
• The restaurant was known for its creative preparation of fish dishes.
• 全体内阁成员祝愿外交大臣的出访取得成功。
• All the cabinet wished the foreign secretary success on her foreign trip.
• 回忆。参与者应该回忆起风、菜肴和旅行。
• RECALL. Participant should recall wind, dishes, and trip.

第一项任务是关注句子并重复最后一个单词。同时，必须调用记忆来存储最后一个单词，以便在最后进行检索。注意这与多任务处理的相似性，即尝试同时执行多个信息处理（Covre、Baddeley、Hitch & Bueno，2019；Popov、Marevic、Rummel & Reder，2019）。这种相似性使我们认识到，工作记忆（即使不一定完全相同）是系统2思维的关键组成部分，它涉及有意识地、策略性地部署资源有限的信息处理。大量研究表明，双跨度测试和其他工作记忆效率指标与流体智力心理测量指标之间的相关性接近1.00（Kyllonen & Christal，1990；Engle et al.，1999；Hunt，2011；Chuderski，2015）。事实上，Conway、Macnamara 和 Engel de Abreu（2013）得出结论：“智力测试可能会被新的工作记忆测试所取代。” 此外，即使工作记忆能力与流体智力并不完全相同，它也已被证明对学业成绩（Alloway 和 Alloway，2010）以及工作记忆受损驾驶员的车祸增加（Walshe 等人，2019）至关重要。

The first task is attending to the sentences and repeating the last word. At the same time, memory must be called on to store the last word for retrieval at the end. Notice the similarity to multitasking, trying to perform more than one information process at the same time (Covre, Baddeley, Hitch, & Bueno, 2019; Popov, Marevic, Rummel, & Reder, 2019). This similarity points us to the idea that working memory, if not necessarily identical to, is the key component of System 2 thinking, involving conscious, strategic, deployment of resource-limited information processes. Extensive research has shown that the correlation between performance on dual span and other measures of working memory efficiency with psychometric measures of fluid intelligence approach 1.00 (Kyllonen, & Christal, 1990; Engle et al., 1999; Hunt, 2011; Chuderski, 2015). Indeed, Conway, Macnamara, & Engel de Abreu (2013) conclude that, “[T]he enterprise of intelligence testing could be replaced by a new enterprise of working memory testing.” Moreover, even if it isn’t identical to fluid intelligence, working memory ability has been shown to be critical to school achievement (Alloway & Alloway, 2010) and to increased crashes in drivers with impaired WM (Walshe et al., 2019).

长期记忆的方面

Aspects of Long-Term Memory

内容可寻址性、要点和推理

Content Addressability, Gist, and Inference

由于LTM的容量是无限的，因此，至关重要的是，其信息必须以易于理解的方式进行组织。在讨论实时问题时，我们以被困在有火灾威胁的房间里为例。在这种情况下，生存取决于快速知道房间里哪些东西能帮助我们逃脱，哪些不能。奥古斯丁关于记忆的直觉性洞穴和洞穴隐喻在此失效，因为它混淆了自传体记忆和知识，而知识正是对有效行为至关重要的记忆储存。

Because the capacity of LTM is infinite, it’s vital that its information be structured in ways that make it accessible. In discussing the real-time problem, we used the example of being caught in a room threatened by fire. In this scenario, survival depended on quickly knowing what items in the room could help us escape, and which could not. Here, Augustine’s intuitive caves and caverns metaphor for memory breaks down because it conflates autobiographical memory and knowledge, which is the memory store important for effective behavior.

与个人记忆不同，语义记忆和知识是内容可寻址的。这意味着有意义相关的想法被存储在一起，因此访问一个想法就可以访问相关的想法。标准计算机内存不是内容可寻址的，构建内容可寻址的存储器一直是人工智能面临的一大挑战。我去英国图书馆的研究之旅体现了内容可寻址和非内容可寻址检索系统之间的重要区别。英国图书馆，例如牛津大学的博德利图书馆，历史悠久，早期的图书管理员只是将书籍送到书架上。为了方便查找，每本书都分配了一个书架号，并逐月逐年地创建了书名和书架号（以及随着新建筑的建造以存储更多书籍，位置也随之变化）的粘贴集合。因此，如果有人在寻找一本特定的书，比如格罗斯的《古物^3》（1772 年），他会找到一本标有大致出版日期的书架号簿，然后翻阅笔记本找到它的位置和书架号。当你到达指定地点时，你的书会放在旁边，旁边是同时送来的其他书。格罗斯的《古物》可能放在修道院史、园艺书、外科书籍或其他任何书籍旁边。至少在当时（1988年），除了少数例外（例如伦敦大学），英国图书馆的书籍是无法进行内容寻址的：书籍的地址（书架号）并不能透露书籍的内容。

Unlike personal memory, semantic memory and knowledge are content addressable. This means that ideas that are meaningfully related are stored together, so that having access to one idea opens up access to related ideas. Standard computer memory is not content addressable and building content addressable memories has been a big challenge to AI. My research trip to British libraries exemplifies the important differences between content addressable and non-content addressable retrieval systems. British libraries, such as Oxford’s Bodleian, are ancient, and in the early days librarians simply placed books on shelves as they arrived. To make books findable, each book was assigned a shelf number, and pasted-together collections of books’ names and shelf numbers (and, as new buildings were built to store yet more books, locations) were created month by month and year by year. Thus, if one were looking for a particular book, say Grose’s Antiquities³ (1772), one would find a shelf number book for the approximate publication date and leaf through the notebook to find its location and shelf number. When you arrive at the location, there would be your book, flanked by other books that came in at the same time. Grose’s Antiquities might be next to a monastic history, a book on gardening, a text on surgery, or anything else. At least at the time (1988) and with a few exceptions (e.g., University of London), books in British libraries were not content addressable: A book’s address (the shelf number) revealed nothing about the content of the book.

你可能熟悉美国图书馆系统，它们的书架是内容可寻址的，要么通过杜威十进制分类法，要么在高等教育中使用国会图书馆（LOC）系统。这里有一本非随机抽样图书的 LOC 书架号：BF81.L4。仅凭地址，你就能了解大部分内容。B 表示你位于图书馆的哲学区。BF 表示你位于哲学区的心理学区，绝大多数心理学书籍都存放在这里。其他一些书架上也有一些，比如自然科学区（Q）、学习区（LB）或精神病学区（RC）。81 表示你位于心理学史和方法论区，可能是历史区，因为它们的编号较低。L 表示作者（谦虚的）姓氏以 L 开头。显然，语义记忆不使用 LOC 或杜威系统来组织其知识库，但它是内容可寻址的。认知科学的经验问题是探索人类记忆如何是如何组织的，以及对于人工智能的工作来说，如何在机器中有效复制，例如专家系统或自动驾驶汽车。

You are probably familiar with American library systems, whose shelves are content addressable, either by the Dewey Decimal System or, in higher education, the Library of Congress (LOC) system. Here’s a LOC shelf number for a book not sampled at random: BF81.L4. From the address alone, you can figure out much of the content. The B tells you you’re in the Philosophy section of the library. BF tells you you’re in the psychology section of Philosophy, where the vast majority of psychology books are shelved. There are a few elsewhere, in Q, natural science, LB, learning, or RC, psychiatry. The 81 tells you you’re in history and methodology of psychology books, probably history because they have the lower numbers. The L tells you that the last name of the (modest) author begins with L. Obviously, semantic memory does not use the LOC or Dewey system to organize its store of knowledge, but it is content addressable. The empirical question for cognitive science is to discover how human memory is organized and, for work in AI, how it can be usably duplicated in machines, such as expert systems or self-driving cars.

这项研究始于Bousfield (1953) 在自由回忆中发现的类别聚类现象。当人们回忆从几个类别中抽取的单词列表时，他们倾向于将同一类别中的单词一起提取出来。长颈鹿、茄子、狒狒的列表就是他提出的一个。两种相互竞争的解释引发了争议。Bousfield认为，参与者首先回忆的是类别，然后是该类别下的项目。另一种观点则更为简单，认为由于这些单词本身已经通过意义关联起来，因此无需援引其在上级类别中的共同成员关系 (Cofer, 1965)。

The search began with Bousfield’s (1953) finding of category clustering in free recall. When people recall lists of words drawn from a few categories, they tend to retrieve items from the same category together. The giraffe, eggplant, baboon list was one of his. Controversy was provoked by two rival explanations. Bousfield thought that participants recalled the categories and then the items under the category. The alternative view was simpler, arguing that because the words themselves were already associated by meaning, invoking common membership in a superordinate category was unnecessary (Cofer, 1965).

工作记忆模型的发展为聚类研究带来了新的启示。聚类并非像联想那样是一个自动的信息处理过程，而可以被视为一种战略决策，即如何编码和检索超出短期记忆（STM）容量的列表，从而对认知资源提出要求。参与者之间存在个体差异。一些人倾向于在语义上聚类他们的回忆，而另一些人则没有，前者比后者记住了列表中的更多单词（Puff、Murphy & Ferrara，1977），这表明前者的信息处理管理效率更高。后来的研究表明，参与者最初如何选择将单词编码为类别成员，以及/或者如何注意到类别并使用它们来帮助在语义长期记忆（LTM）中搜索列表中的成员，这支持了以下观点：类别列表学习是对工作记忆的双重任务评估，任务包括项目存储和检索，以及注意并记住未提及的类别标签（Gruenewald & Lockhead，1980；Cinan，2003）。

The development of models of working memory threw new light on clustering. Clustering was not an automatic information process as was association but could be viewed as a strategic decision about how to encode and retrieve a list that exceeded the capacity of STM and thus placed demands on cognitive resources. There were individual differences between participants. Some tended to semantically cluster their recalls and others did not, with the former remembering more words from the list than the latter (Puff, Murphy, & Ferrara, 1977), suggesting more efficient management of information processing. Later studies showed that how participants chose to initially encode words as members of categories and/or noticed the categories and used them to help search semantic LTM for the members that were on the list, supported the idea that categorical list learning was a dual task assessment of working memory, the tasks being item storage and retrieval plus noticing and keeping in mind the unmentioned category labels (Gruenewald & Lockhead, 1980; Cinan, 2003).

Deese (1959)、Roediger 和 McDermott (1995、2000) 范式或 DRM 范式说明了语义联想对记忆的强大影响，并且它为后来的研究主题——虚假记忆的产生——提供了启示（Gallo, 2010）。与 Bousefield 范式一样，参与者听到或阅读一串他们之后会回忆起来的单词。以下是其中一个 DRM 单词表：休息、睡觉、打瞌睡、疲倦、打鼾、醒来、小睡、清醒、做梦、平静、瞌睡、打盹、打哈欠、毯子、沉睡。在此单词表中，所有单词都是相关的，因此我们看不到任何聚集现象。然而，我们确实发现了一个不包含在列表中的“回忆起的”单词，即诱导词 SLEEP。由于 sleep 与列表中所有其他单词都有关联，因此接触该单词表就产生了虚假记忆。

The Deese (1959), Roediger, McDermott (1995, 2000), or DRM, paradigm illustrates the powerful effect of semantic association on memory, and it throws light on a later topic, the creation of false memories (Gallo, 2010). As in the Bousefield paradigm, participants hear or read a list of words that they later recall. Here is one DRM list: REST BED DOZE TIRED SNORE WAKE NAP AWAKE DREAM PEACE DROWSY SNOOZE YAWN BLANKET SLUMBER. In this list, all the words are related, so we don’t see any clustering. What we do find, however, is a “recalled” word not contained in the list, the lure word, SLEEP. Because sleep is associated with all the other words on the list, exposure to the list has created a false memory.

然而，长时记忆的结构远不止语义联想，正如舌尖现象（简称TOT）所体现的那样。每个人都经历过这种意识状态——几乎但并非完全能够回忆起一个单词或名字——但第一位对此进行评论的心理学家是威廉·詹姆斯（1890）。詹姆斯指出，TOT状态是意识中的一种缺口，但这种缺口非常活跃。正如奥古斯丁所观察到的，如果朋友提出答案，我们会拒绝错误的答案；我们会感到焦虑，只有当正确的词语最终出现时，我们才会感到轻松。正是这种活动使TOT成为奥古斯丁记忆悖论的一个绝佳例证。显然，从某种意义上说，我们知道我们在记忆中寻找，但同样明显的是，我们并没有。在认知科学教师讨论的所有认知现象中，这是我们最常被问到的问题：为什么会发生TOT？我们能做些什么来应对？

However, there’s more to the structure of LTM than semantic association, as shown by the tip-of-the-tongue phenomenon, or TOT for short. Everybody has experienced this state of consciousness—almost but not quite being able to retrieve a word or name—but the first psychologist to remark upon it was William James (1890). James pointed out that the TOT state was a gap in consciousness, but it was a very active gap. As Augustine observed, if friends suggest answers, we reject wrong ones; we feel anxious, and relief comes only when the right word finally emerges. It’s the activity that makes TOT a wonderful example of Augustine’s paradox of memory. Clearly, in one sense we know what we’re looking for in memory, but equally clearly, we don’t. Of all the cognitive phenomena teachers of cognitive science discuss, it’s the one we’re most often asked about: Why does TOT happen and what can one do about it?

不幸的是，这很难进行研究，因为我们无法随心所欲地让它发生。布朗和麦克尼尔（1966）开发了一种程序，只要有足够多的参与者，通常至少会在少数人身上引发TOT。他们召集了几百名参与者，向他们宣读他们希望出现在部分参与者词汇表中但数量又足够少的单词的定义，以至于他们可能会陷入TOT。我用这个方法在课堂上进行了35年的演示，发现有时效果很好，有时却惨不忍睹。

Unfortunately, it’s very hard to investigate because we can’t make it happen at will. Brown and McNeill (1966) developed a procedure that, given enough participants, will usually induce TOT in at least a few. They gathered several hundred participants together and read them definitions of words that they hoped would be in some participants’ vocabularies but were rare enough that they might fall into TOT. I used this as a class demonstration for over 35 years and found that sometimes it worked beautifully and sometimes it failed miserably.

我将以我遇到这个方法时进入TOT状态的单词为例。它的定义如下。如果幸运的话，你就能体验到TOT。

I’ll use as an example the word that put me in TOT when I met the method. Here’s the definition. With luck, you will experience TOT.

一种具有 60° 刻度弧和瞄准装置的仪器，用于测量物体之间的角距离，尤其是用于在导航中测量高度。

An instrument with a graduated arc of 60° and a sighting mechanism, used for measuring the angular distances between objects and especially for taking altitudes in navigation.

报告称处于 TOT 状态的参与者随后填写了一份简短的问卷，例如询问“你能说出第一个音节吗？”“你能想到押韵的吗？”“你能想到一个意思相近的词吗？”

Participants who reported being in TOT then filled out a brief questionnaire, asking, for example, “Can you name the first syllable?” “Can you think of a rhyme?” “Can you think of a word of similar meaning?”

如果参与者体验到TOT状态，他们就能像在现实生活中一样，回答其中一些问题，并拒绝关于目标词的错误建议，而无需提取物品本身。就我而言，作为一名历史爱好者，我毫不意外地得到了该物品的中世纪祖先——星盘，而没有得到六分仪本身（图6.4）。在我的一堂课上，一个学生可以画出一个六分仪，而无需说出它的名字。

If they experience the TOT state, participants can, as in real-life TOT, answer some of these questions and reject wrong suggestions about the target word, without retrieving the item itself. In my case, and unsurprisingly for a history buff, I got the item’s medieval ancestor, the astrolabe, without getting the item itself, the sextant (Figure 6.4). In one of my classes, a student could draw a sextant without naming it.

尽管对研究提出了挑战，但TOT状态本身就足够引人入胜，并已引发了大量研究（Brown，1991，2012）。然而，由于其脆弱性和难以捉摸的性质，无论是在理论层面还是实践层面，都鲜有探讨。两位研究者（Taylor & Mackay，2003）在评论一本关于TOT的书籍时，称其既是一座金矿，也是一座充满挑战的宝库。

Despite the challenge it presents to investigation, the TOT state is fascinating enough to have generated a great deal of research (Brown, 1991, 2012). Unfortunately, because of its fragility and elusive nature, there is not much to say about it theoretically or practically. Reviewing a book on TOT, two investigators (Taylor & Mackay, 2003) call it both a gold mine of findings and a can of worms.

就我们的目的而言，TOT 状态表明，词语并非仅仅通过其含义存储在 LTM 中（正如类别聚类和 DRM 结果所暗示的那样）。相反，词语在记忆中也被编码为具有特征性声音和音节划分的物理对象，并且通常附带图像甚至历史故事（星盘看起来不像六分仪）。在检索时，必须具备一定数量的最低限度的特征才能使整个词语浮现在脑海中。

For our purposes, the TOT state shows that words are not stored in LTM solely by their meanings, as category clustering and the DRM results might suggest. Instead, words are also encoded in memory as physical objects with characteristic sounds and syllabic divisions, and often with images and even historical stories attached (an astrolabe does not look like a sextant). At retrieval, some minimum number of characteristics must be available for the whole word to come to mind.

最后，句子记忆表明，记忆行为更多的是重构而非记忆。正如普鲁斯特所写：“回忆过去的事物不一定是回忆事物的本来面目。” 威廉·F·布鲁尔突破了以词汇为中心的语言学习传统，进行了一系列关于句子记忆的长期实验，揭示了呈现的句子在编码过程中如何系统地发生改变。我（他的研究助理）阅读了大约十几个句子的列表，然后进行了一些简单的心算（例如，32 × 74 的乘法），以消除近因效应。然后，参与者收到一张写有每个句子前几个单词的答题纸，并被要求完成每个句子。

Finally, memory for sentences shows that the act of remembering is more reconstruction than memory. As Proust wrote, “Remembrance of things past is not necessarily remembrance of things as they were.” Moving past the word-centric verbal learning tradition, William F. Brewer conducted a long series of experiments on sentence memory, demonstrating systematic ways in which presented sentences were altered during encoding. I (his research assistant) read lists of about a dozen sentences, which were followed by performance of some simple mental algebra (e.g., multiply 32 × 74) to eliminate recency effects. Then, participants were given an answer sheet with the first few words of each sentence and were asked simply to complete each one.

其中一个句子是“子弹击中了靶心”。在答题纸上，参与者填写的是“子弹 _____________________”。绝大多数学生回忆起的是“子弹击中了靶心”（Brewer，1975）。他提出，参与者并没有存储包含“struck”一词的句子的表面结构，而是存储了句子的整体含义，传统上称为 gist 。在检索时，gist 被正确地记住了，但被包装成了“struck”的同义词“hit”，这在口语英语中更为常见。gist 的概念也在 DRM 范式的一个理论中得到运用。从某种意义上说，bed、rest、……等单词列表的 gist 与 sleep 有关，因此参与者错误地回忆起它曾在原始列表中（Gallo，2010）。

One sentence was, “The bullet struck the bullseye.” On the answer sheet, participants completed, “The bullet _____________________.” Students overwhelmingly recalled, “The bullet hit the bullseye” (Brewer, 1975). He proposed that participants did not store the surface structure of the sentence with the word “struck,” but they stored the overall meaning of the sentence, traditionally called the gist. At retrieval, the gist was remembered correctly but dressed up in a synonym for “struck,” “hit,” which is a little more common in colloquial English. The idea of gist is also deployed in one theory of the DRM paradigm. In a sense, the gist of the list of words bed, rest, … etc., concerns sleep, so participants falsely recall it as having been on the original list (Gallo, 2010).

更复杂的处理过程可以更显著地改变回忆，赋予句子一个隐含的背景故事。例如，“心不在焉的教授没有车钥匙”几乎从未被准确回忆，而是以某种形式被回忆成“心不在焉的教授忘记了[或丢失了]他的车钥匙”（Brewer，1977）。在这种情况下，参与者推断出钥匙丢失的合理原因，并将其包含在他们之后回忆的句子主旨中。记忆是对经验的合理重构而非记录，这一观点对于理解Simons和Chabris（2011）记录的一些记忆迷思至关重要。

More complex processing can change recall more dramatically, giving sentences an implicit backstory. For example, “The absent-minded professor didn’t have his car keys” is almost never recalled accurately, but as some form of “The absent-minded professor forgot [or had lost] his car keys” (Brewer, 1977). In this case, participants inferred a plausible reason for the missing keys and included it in the stored gist of the sentence that they recalled later. The idea that memory is a plausible reconstruction of experience rather than its recording is important to understanding some of the myths of memory documented by Simons and Chabris (2011).

到目前为止，我们关注的研究主要集中在记忆在提取时是如何形成的，尽管我们也注意到了提取过程中的编码过程。然而，有两种理论观点强调了编码在决定记忆内容方面的作用。

The research we have looked at so far has focused on how memories are shaped at the time of retrieval, although we have noted encoding processes along the way. Two theoretical perspectives, however, emphasize the role of encoding in determining what is remembered.

记忆作为想象的建构和重构：图式

Memory as Imaginative Construction and Reconstruction: Schemas

到目前为止，除了一个例外，我们讨论记忆的理论都是基于一种直觉理论，即痕迹理论（Bartlett，1932；Wagoner，2013）。该理论认为，经历会在我们的大脑/心灵中留下痕迹，这些痕迹可以被激活，记忆就能被准确而完整地提取出来。普鲁斯特对逝去时光的追寻，出乎意料地始于一天，那天他喝了茶，吃了一块他年轻时吃过的特别饼干。他陷入了一段快乐却漫无目的的遐想，试图在几天的茶和饼干中重温这段遐想，并思考自己的感受从何而来。最终，茶和饼干的味道奏效了——他回到了在贡布雷的童年时光，回忆起教堂、街道、学校，事实上，“贡布雷及其周边环境，无论是城镇还是花园，都从我的那杯茶中，逐渐成形，逐渐成形。”他对贡布雷的记忆痕迹因久违而消逝，但饼干和茶解开了一切的谜团。

So far, with one exception, we have discussed memory in terms of an intuitive theory called trace theory (Bartlett, 1932; Wagoner, 2013). The idea is that experiences leave traces behind in our brain/mind, that these traces can be activated, and the memories retrieved accurately and in toto. Proust’s search for lost time unexpectedly begins one day when he drinks tea and eats a special cookie from his youth. He enters a happy but unfocused reverie that he tries to recapture over several days of tea and cookies, wondering whence came his feelings. Finally, the taste of the tea and cookie works—he’s thrown back into his childhood in Combray, remembering the church, the streets, the school, indeed “the whole of Combray and its surroundings, taking shape and solidity, sprang into being, town and gardens alike, from my cup of tea.” The traces of his memory of Combray had become lost with disuse, but the cookies and tea unlocked everything.

记忆痕迹理论最早由联想主义者（Brewer & Nakamura，1984）正式提出，并由此通过艾宾浩斯的《论记忆》进入心理学，并延续到言语学习研究中。记忆中的信息以各种方式编码和关联，它们之间的联想联系决定了哪些信息可以被提取或遗忘。

Trace theory of memory had been more formally stated by associationists (Brewer & Nakamura, 1984), and from them it entered psychology via Ebbinghaus’s On Memory, and continued in verbal learning research. Items in memory are encoded and associated in various ways, and the associative links between them determine what can be retrieved or forgotten.

我们对痕迹理论的一个有益的例外是记忆中的句子“心不在焉的教授没有他的车钥匙”，这句话通常被误读为“心不在焉的教授忘记了[或丢失了]他的车钥匙”。诚然，这里有一个痕迹，但记忆已经被一个既不是遗忘（没有钥匙的基本事实仍然存在）也不是联想效应的过程所改变——“没有钥匙”与“忘记”或“丢失”没有关联。相反，对这句话的错误记忆是由于泛泛知识，即那些并非关于某个人、某个地点或某件事，而是关于一个多面刻板印象的知识。我们得知自己面对的是一位心不在焉的教授，于是自然而然地会用这个刻板印象来解释他为什么没有带车钥匙。

Our instructive exception to trace theory was memory for the sentence, “The absent minded-professor didn’t have his car keys,” which is typically mis-retrieved as “The absent-minded professor forgot [or lost] his car keys.” There is a trace here, to be sure, but the memory has been altered by a process that is neither forgetting—the basic fact of not having keys remains—nor an effect of association—didn’t have has no association with forgot or lost. Instead, the misremembering of this sentence is due to a piece of generic knowledge, that is, knowledge that isn’t about a person, place, or thing, but about a multifaceted stereotype. We learn that we’re dealing with an absent-minded professor, and so we naturally use the stereotype to provide a reason why he doesn’t have his car keys.

我们拥有许多高度结构化的通用知识。你对新的诊所是什么样子以及在那里会发生什么有着预期。你对校园有一个通用的心理地图；不是记忆中的图像，而是一幅缺乏情景细节（例如天气）的图画。你有一个关于如何在各种餐厅行动的脚本（Shank & Abelson，1977）；当你走进一家新餐厅时，你会寻找线索来判断它是哪种类型的。有桌布吗？那么你应该等待入座，即使没有任何指示。所有这些都是认知科学中最重要的概念之一的一个例子，即图式，它是由英国心理学家弗雷德里克·巴特利特 (Bartlett, 1932; Brewer, 2000; Wagoner, 2013) 首次提出的，尽管早些时候冯特 (Wundt) 以Gesamtvorstellung 的术语提出过，即“有组织的整体思想”(Leahey, 2018; Leahey & Holtzman, 1979)。

We possess many different highly structured pieces of generic knowledge. You have expectations about what a new doctor’s office will look like and what will happen there. You have a generic mental map of your campus; not an image from memory, but a picture lacking episodic detail, such as weather. You have a script (Shank & Abelson, 1977) for how to act in a variety of restaurants; when you enter a new restaurant, you look for cues as to which type it is. Are there tablecloths? Then you should wait to be seated even if there’s no sign to tell you so. All of these are examples of one of the most important concepts in cognitive science, the schema, first proposed by British psychologist Frederick Bartlett (Bartlett, 1932; Brewer, 2000; Wagoner, 2013), although earlier offered by Wundt under the term Gesamtvorstellung, “organized whole idea” (Leahey, 2018; Leahey & Holtzman, 1979).

巴特利特最著名的研究是《幽灵之战》。巴特利特认为，痕迹理论从根本上来说存在缺陷，因为它将记忆视为一种被动记录，由感觉刻蚀笔在蜡板上形成，在适当的情况下可读。相反，巴特利特认为，记忆是通过人类在追寻意义后所进行的主动推理过程（重新）建构的。我们不是经验的被动记录者，而是主动构建对经验的最佳诠释，借鉴一生中积累的想法、概念、期望和态度，然后将它们作为动态心理结构（称为图式）存储在记忆中。当需要检索记忆时，建构过程会再次发生，我们会创建最可信的过去画面。我们不知道最初的建构过程、后来的重建过程，也不知道它们所利用的图式，这给心理学研究带来了挑战。如何揭示建构和重建过程？巴特利特的《幽灵之战》试图应对这一挑战。

Bartlett’s most famous study involved The War of the Ghosts. Bartlett believed that trace theory was fundamentally flawed because it saw memory as a passive recorder, created by the stylus of sensation etching on a wax tablet, readable under the right circumstances. Instead, Bartlett argued, memories were (re)constructed by an active process of inferences guided by human effort after meaning. Rather than being passive recorders of experience, we actively construct the best interpretation of experience we can, drawing on ideas, concepts, expectations, and attitudes built up over a lifetime, and these are then stored in memory as dynamic mental structures called schemas. When it comes time to retrieve a memory, the constructive process occurs again as we create the most plausible pictures of our pasts. We are not aware of the process of the original process of construction, the later process of reconstruction, nor of the schemas they exploit, creating a challenge for psychological research. How can the construction and reconstruction processes be revealed? Bartlett’s War of the Ghosts attempted to meet the challenge.

文化创造图式。一个没有餐馆或餐馆出现之前的文化，不会有各种各样的餐馆脚本。一个没有巫师的文化，不会有关于巫师的图式。只有狩猎成为日常活动的文化，才会有关于狩猎的图式。不同文化（和时代）对来世以及鬼魂的存在和行为的图式也大相径庭。因此，巴特利特选择了一个加拿大本土文化的民间故事，并要求他的英国参与者聆听并回忆这个故事。《鬼魂之战》中默默嵌入的图式与英国文化图式并不相符，因此，当他的参与者试图从他们所听到的内容中寻找意义时，他们会运用自己的图式，将其揭示为记忆的扭曲。我开展了巴特利特的研究长达35年，研究了数千名学生，我将与你们分享我最喜欢的一段回忆（图6.5）。它的独特之处在于它以绘画而非口头记忆的形式呈现，但它是重建记忆的一个绝佳典型例子。

Cultures create schemas. A culture without, or in an era before, restaurants, will not have varied restaurant scripts. A culture without shamans will not have schemas about them. Only a culture in which hunting is routine will have schemas about hunting. And cultures (and eras) differ greatly in their schemas about the afterlife and the existence and behavior of ghosts. So, Bartlett chose a folktale from a Canadian native culture and asked his British participants to listen to and recall it. The schemas silently embedded in the War of the Ghosts would not match up with British cultural schemas, so that as his participants tried to find meaning in what they heard, they would apply their own schemas, revealing them as distortions of memory. I ran Bartlett’s study for 35 years and with thousands of students, and I will share with you my all-time favorite recall (Figure 6.5). It’s unique in being presented as a drawing rather than as a verbal memory, but it’s a wonderfully typical example of reconstructive memory.

这里有迹可循：两个年轻人、战士、独木舟、战斗、篝火旁的村民。然而，许多细节已丢失；比如为什么这两个来自埃古拉克的年轻人会在河边。在这种情况下，用图画说明原因可能比较困难，这也凸显了巴特利特创造性但非结构化和开放式程序的缺陷。然而，在书面回忆中，他们晚上外出的原因——猎海豹——即使在第一次直接回忆中也常常是开放的。通常，它会被替换为钓鱼，甚至是钓海豹，尤其是在第二次回忆中，然后就变成了钓鱼。巴特利特的英国参与者没有猎海豹的图式，所以他们失去了基本原理，或者用更英国的钓鱼图式取而代之。

There are traces here: The two young men, the warriors, the canoes, the battle, the villagers around the fire. Many details, however, have been lost; such as why the two young men from Egulac were down near the river. It could be that in this case showing the cause might be hard in a drawing, which highlights shortcomings of Bartlett’s creative but unstructured and open-ended procedure. However, in written recalls the reason for their evening out—seal hunting—is often left open even in the first, immediate recall. Often, it’s replaced by fishing, or even by fishing for seals, especially in a second recall before becoming simply fishing. Bartlett’s British participants didn’t have a schema for seal hunting, and so they lost the rationale or replaced it with the more British schema of fishing.

然而，在这个样本中，最能揭示重建记忆的变化在于参战年轻人的命运以及对鬼魂主题的处理。故事中，年轻人中枪了，但他并没有感到不适，直到第二天早上他向村子讲述了自己的故事后才死去。其他战士是鬼魂，尽管没有解释为什么或年轻人是如何发现这一点的。在回忆录中，年轻人在战斗中阵亡，尸体被送回村子。最终，这位年轻的英雄并没有像故事中那样死去，而是复活了。因此，无法解释的鬼魂和一个令人费解的死亡，对于加拿大本土人来说一定是具有图式意义的，但却被强大的欧洲死亡和变形图式所取代，让 21 世纪的美国人民更能理解这个故事。

In this specimen, however, the changes that best reveal reconstructive memory concern the fate of the young man who goes to war and the handling of the ghost theme. In the story, the young man is shot, but he does not feel ill and he does not die until the morning after he tells his tale to the village. The other warriors were the ghosts, though no explanation of why or how the young man figures it out is given. In the recall, the young man is killed in battle and his body is returned to the village. In the end, instead of dying, as in the story, the young hero is resurrected. Thus, unexplained ghosts and a puzzling death, which must have made schematic sense to the native Canadians, are replaced by a powerful European schema of death and transfiguration, making more sense of the story for a 21st-century American.

巴特利特的思想和发现颇具启发性，但由于种种原因，包括行为主义在美国的主导地位以及联想主义在世界范围内的盛行，这些思想和发现影响力甚微，巴特利特的传记作者甚至宣告图式理论已死。然而，随着认知科学（包括重要的人工智能）在20世纪70年代开始研究更高级的心理过程，行为主义和联想主义的缺陷逐渐显现。在探索处理思维，尤其是专业知识的方法的过程中，认知科学家重新发现了图式（Brewer & Nakamura，1984）。

Bartlett’s ideas and findings were suggestive, but for a variety of reasons, including the dominance of behaviorism in America and of associationism everywhere, they exerted little influence and the schema idea was pronounced dead by Bartlett’s biographers. However, as cognitive science, including importantly AI, began to study the higher mental processes in the 1970s, the shortcomings of behaviorism and associationism became apparent. Looking around for ways to handle thinking and especially expertise, cognitive scientists rediscovered the schema (Brewer & Nakamura, 1984).

然而，巴特利特略显模糊和整体化的图式版本需要理论上的完善，而他印象派的准实验结果则需要重复和更定量的处理。在理论方面，明斯基（1975）发表了一篇里程碑式的论文。明斯基致力于人工智能研究，其中的一个主要问题是：如何以清晰易懂的方式表示复杂的概念及其之间的关系，从而使人工智能能够模拟人类智能。他以一个新名称——框架——重新定义了图式的概念，并对其进行了更精确的定义，以便在人工智能系统中部署。

However, Bartlett’s somewhat vague and holistic version of the schema needed theoretical sharpening, and his impressionistic quasi-experimental results needed replication and more quantitative treatment. On the theoretical side, a landmark paper was Minsky (1975). Minsky worked in AI, wherein how to represent complex concepts and relations between them in perspicuous ways so that AI could emulate human intelligence was a major issue. He revived the concept of schema under a new name, the frame, which he defined more precisely for deployment in AI systems.

框架是一种数据结构，由连接到一组槽位的节点连接在一起，每个槽位都包含用于存储框架主体信息的位置。槽位提供了用于编码通常在刻板印象情境中遇到的信息的位置。具有槽位的体验特征很容易被注意到和回忆，而与图式不一致的特征通常会被忽略或无法回忆，就像隐形大猩猩的例子一样。当你的球赛图式被激活时，你往往看不到闯入的大猩猩。一些槽位已经填充了默认值，这些默认值反映了在相关事件中通常会发现的特征；因此，它们通常会覆盖实际情景记忆的编码，而被记住。此外，随着记忆随时间推移而变化，即使情景事件最初被编码，变化也会逐渐被默认值取代。一些默认设置具有反映不同刻板印象设置的范围值，例如作为一名员工或一个家庭男人，因为同一个人可以被编码到不同的图式中。这标志着图式/框架的一个重要方面。它们并非孤立存在，而是构成一个由相互关联的模式组成的整体网络的一部分。例如，医生办公室的模式会包含一个指向其他模式（例如医生、药品、建筑物等）的指针，以及指向模式脚本（例如医生办公室内发生的事情等）的指针^。4

A frame is a data structure held together by a node that is connected to a set of slots, each of which contains places to store information about the subject of the frame. Slots provide locations in which to encode information typically encountered in stereotyped situations. Features of an experience that have slots will be easily noticed and recalled, whereas schema incongruent features will often be missed or not recalled, as in the case of the invisible gorilla. When your ball-game schema is activated, you tend not to see the intruding gorilla. Some slots are already filled in with default values reflecting features typically found in the event in question; thus, they will often override the encoding of actual episodic memories and be remembered instead. Also, as memories change over time, even when episodic events were initially encoded, change will drift toward replacement by the defaults. Some of the default settings have range values reflecting different stereotyped settings, such as being an employee or a family man, as the same person can be encoded in different schemas. This marks an important aspect of schemas/frames. They do not exist as isolated entities but are part of a holistic web of interconnected schemas. A doctor’s office schema, for example, will have a pointer to other schemas such as physician, medicine, building, etc., and to schema-scripts such what happens in a doctor’s office, etc.⁴

当你形成记忆时，图式会被实例化（Rumelhart，1980）。当你走进医生的办公室，或者遇到一位J. Smith先生时，你开始形成关于……的情景记忆。事件本身。根据图式理论，你激活了一个或多个图式，它们会引导你对事件的感知、解读和编码，并在之后引导你对事件的回忆。简而言之，你会用情景信息填充框架的槽位，并遵循我们之前提到的编码约束和过程。存储在记忆中的是填充好的、实例化的图式，而这个图式在存储过程中可能会进一步改变，并在回忆时被重新提取。

When you form a memory, schemas are instantiated (Rumelhart, 1980). When you enter a doctor’s office or meet a Mr. J. Smith, you begin to form an episodic memory of the event. According to schema theory, you have activated one or more schemas, and they guide your perception, interpretation, and encoding of the event, and will later guide your recall of it. In short, you fill in the frame’s slots with episodic information, subject to the sort of encoding constraints and processes we’ve noted. What gets stored in memory is the filled-in, instantiated schema, and it is this schema that may be further altered during storage and will be retrieved at recall.

在一项被广泛引用的研究中，Brewer 和 Treyens (1981) 在实验控制下，利用自然场景来检验图式理论的预测。参与者报告了研究结果，但被告知实验室尚未准备好，于是在他们误认为是研究生助理办公室的地方（图 6.6）等候了几分钟。之后，他们被带到另一个房间，根据实验条件，通过绘画、描述或进行再认记忆测试（将房间里曾经存在和不存在的物品混在一起）来回忆起办公室。

In a widely cited study, Brewer and Treyens (1981) used a naturalistic setting under experimental control to test predictions from schema theory. Participants reported for a study but were told that the lab was not quite ready and waited for a few minutes in what they believed to be the graduate research assistant’s office (Figure 6.6). They were then taken to another room, where, depending on experimental condition, they recalled the office by drawing it, describing it, or performing a recognition memory test, mixing up items that had been in the room and items that had not.

仔细观察照片，你会发现有些物品符合研究生的刻板印象，比如桌椅、打字机（当时的）或海报。其他一些物品则出乎意料，比如骷髅或酒瓶。还有很多物品则介于两者之间，比如雨伞或螺丝刀。重要的是，照片中没有典型的研究生办公用品；没有书籍或期刊。

Inspection of the photo will show that some of the items are stereotypical for a grad student schema/frame, such as the tables and chairs, the typewriter (at the time), or the poster. Others are unexpected, such as the skull or the wine bottle. Many are in between, such as the umbrella or the screwdriver. Importantly, typical grad student office items are absent; there are no books or journals.

除了一个例外，所有回忆方法的结果都支持了 Brewer 和 Treyen 基于图式理论的预测。符合图式预期的物品（例如打字机）尤其容易被记住。符合图式预期的物品预期但实际缺失的事物（例如书籍）经常被错误地报告为存在。他们增加了一个新的预测，将图式理论与语言学中关于对话的“既定-新”契约相结合。随着对话的进行，对话会积累知识，每个人都将已经说过的内容视为既定内容，并只添加新的想法。在自由回忆条件下，Brewer 和 Treyens 预测图式会提供大量无需重复的既定事实。因此，所有房间都有墙壁、地板和天花板，因此它们都是更大的房间或办公室框架的一部分，可以被视为理所当然，即使它们是可以编码的情景数据的一部分，正如在再认测试中发现的那样。有趣的是，一些参与者在自由回忆条件下确实报告了墙壁等信息^。5

With one exception, the results from all recall methods supported Brewer and Treyen’s predictions from schema theory. Present items that matched schema expectations (e.g., the typewriter) were especially well remembered. Items that fit schema expectations but were absent (e.g., books) were often falsely reported as present. They added a new prediction that combines schema theory with what is known in linguistics as the given-new contract about conversations. As they proceed, conversations accumulate knowledge, each person taking as given what’s already said and adding only new ideas. In the free-recall conditions, Brewer and Treyens predicted that schemas provide a large number of given facts that don’t need to be repeated. Thus, all rooms have walls, a floor, and a ceiling, and so are part of a larger room or office frame that can be taken for granted, even though they are part of the episodic data that could be encoded, as found during recognition tests. A few participants did, interestingly, report the walls and so on in the free recall conditions.⁵

意外的发现是，与研究生图式明显不同的项目（例如头骨）被很好地记住，甚至比图式一致的项目更容易被记住。Pezdek 等人 (1989) 在后续研究中探究了原因，该研究重复了 Brewer 和 Treyens 的实验。Pezdek 等人提出，由于差异项目不符合图式，因此它们在工作记忆中会得到更多的处理。此外，差异项目的查看时间更长，因此它们比预期项目编码得更详细，使它们以后更容易记住。考虑到我们对工作记忆和记忆监控的讨论，这些提议可以结合起来说，WM 中的执行过程会将更多的时间和注意力分配给令人惊讶的输入（与预测处理架构中的错误传播消息一致），从而使意外项目更重要，因此更有可能被存储和检索。

The unexpected finding was that items saliently discrepant from the grad student schema (e.g., the skull) were well remembered, even better remembered than schema-consistent items. Why was explored in a follow-up study by Pezdek et al. (1989) that replicated Brewer and Treyens’s experiment. Pezdek et al. proposed that because discrepant items don’t fit the schema they receive more processing in working memory. Additionally, discrepant items are looked at longer, and thus they are encoded in more detail than expected items, making them easier to remember later. Considering our discussion of working memory and memory monitoring, these proposals can be combined by saying that executive processes in WM allocate more time and attention to surprising inputs (consistent with error propagation messages in the predictive processing architecture), making unexpected items more important and therefore more likely to be stored and retrieved.

记忆是一种心理时间旅行，从现在回到过去。图式在心理时间旅行中也发挥着作用（Schacter、Addis & Buckner，2007），即想象我们可能做的事情或成就，例如赢得诺贝尔奖或成为美国总统或英国首相。如果说记忆是一种想象性的重构，正如巴特利特所言，那么想象未来也是如此，因为可能发生的事件是由图式构建的。鲁宾（Rubin，2014）要求参与者回忆过去并想象未来的痛苦事件，结果发现未来事件被评为比过去事件更具创伤性。事实上，未来事件被评为非常负面，以至于符合创伤后应激障碍（PTSD）的诊断标准。同样，哈特（Hart）、范·埃普斯（van Epps）和施韦策（Schweitzer，2019）发现，人们在谈话中通常不会问敏感问题，因为他们认为这样做的后果比实际更糟糕。

Remembering is a sort of mental time travel, moving from the present to the past. Schemas also play a role in mental time travel to the future (Schacter, Addis, & Buckner, 2007), imagining things we might do or achieve such as winning the Nobel Prize or becoming US president or British prime minister. If memory is an imaginative reconstruction, as Bartlett suggested, then imagining the future is, too, as possible events are constructed from schematic ones. Asking participants to recall past and to imagine future painful events, Rubin (2014) found that future events were rated as more traumatic than past ones. In fact, future events were rated so negatively that they met the diagnostic criteria for post-traumatic stress disorder (PTSD). Similarly, Hart, van Epps, and Schweitzer (2019) found that people often don’t ask sensitive questions during conversations because they imagine the consequences of doing so to be worse than they would be.

记忆结构在专业知识中起着关键作用。临床心理学家被认为是人类行为方面的专家，但他们的专业知识却是由相互关联的图式构成的 (Kim & Ahn, 2002)。人们或许认为，现实世界的经验能够提升从课堂和书籍中习得的临床判断能力，但事实并非如此。临床医生的经验越丰富，就越有可能记错病例，从而将过于明确的结构强加于患者身上 (Weine & Kim, 2018)。

Memory structures play a key role in expertise. Clinical psychologists are supposed to be experts in human behavior, but their expertise is structured as patterns of interlocking schemas (Kim & Ahn, 2002). One might think that real-world experience would improve clinical judgment learned from classes and books, but it turns out not to be true. The more experience a clinician has, the more likely it is that he or she will misremember cases, imposing more clear-cut structure on them than is warranted (Weine & Kim, 2018).

最后，顺着明斯基的说法，我一直把图式/框架称为“刻板印象”。我们所认为的刻板印象，其实只是一套关于我们社会世界的图式。我们几乎对万事万物都有图式。鉴于“刻板印象”一词如今带有污名，你可能会怀疑它们是否准确。关于哺乳动物神经系统的一个事实是，它非常擅长在自身经验中发现并响应模式。事实上，它最基本的认知功能很可能就是良好的模式发现和物体分类能力，这种能力在数百万年的进化过程中得到了磨练和提升。因此，即使人类社会生活是进化的新生事物，发现模式和形成图式（也就是刻板印象）的能力应该仍然有效。在多年的一系列研究中，Lee Jussim 和他的同事表明“刻板印象的准确性是社会心理学中最大和最可重复的发现之一”（Jussim、Crawford & Rubenstein，2015 年，第 490 页；Jussim、Cain、Crawford、Harber & Cohen，2009 年；Jussim、Crawford、Anglin、Chambers、Stevens & Cohen，2016 年；Jussim、Stevens & Honeycutt，2018 年）。

Finally, following Minsky, I have been calling schemas/frames “stereotypes.” What we think of as stereotypes are just a set of schemas about our social world. We have schemas for just about everything. Given that the term “stereotype” carries an aura of stigma today, you might wonder if they are accurate. One fact about the mammalian nervous system is that it’s very good at finding and responding to patterns in its experience. In fact, its most basic cognitive function is probably good pattern-finding and object classification, honed and improved over millions of years of evolution. Therefore, even though human social life is an evolutionary newcomer, finding patterns and forming schemas—stereotypes—should still work well. In a series of studies over many years, Lee Jussim and his colleagues have shown that “stereotype accuracy is one of the largest and most replicable findings in social psychology” (Jussim, Crawford, & Rubenstein, 2015, p. 490; Jussim, Cain, Crawford, Harber, & Cohen, 2009; Jussim, Crawford, Anglin, Chambers, Stevens, & Cohen, 2016; Jussim, Stevens, & Honeycutt, 2018).

关于长期记忆的争议

Controversies about Long-Term Memory

闪光灯记忆：记忆的生动性是否表明准确性？

Flashbulb Memories: Does Vividness of Memory Indicate Accuracy?

Brown & Kulik (1977) 提出了“闪光灯记忆”的概念，即对令人意外的情绪事件的生动且看似准确的记忆。有趣的是，他们并没有真正断言这些记忆是准确的（Brewer，2006），而是将其视为理所当然。他们试图通过采访人们关于他们在1963年11月22日得知约翰·F·肯尼迪遇刺事件时的记忆来描述这一新提出的记忆类型^。7

Brown & Kulik (1977) introduced the idea of “flashbulb” memories, vivid and seemingly accurate memories for surprising emotional events. Interestingly, they did not actually assert that the memories were accurate (Brewer, 2006) but took it for granted. They attempted to characterize this new proposed class of memories by interviewing people about their memories of learning of John F. Kennedy’s assassination on November 22, 1963.⁷

从方法论上讲，闪光灯记忆的研究非常困难。像布朗和库利克这样的回顾性研究无法解决准确性问题，因为没有参与者原始记忆的记录作为基准，也没有同期的“普通”记忆作为对照，来比较闪光灯记忆的生动性和准确性。当然，科学家不可能仅仅为了做实验就引发惊天动地的事件。逐渐形成的一个现成的研究方案是，在闪光灯事件发生时，手头上要准备好调查问卷，以便快速采访参与者（通常是心理学课程的学生），了解他们对闪光灯事件和对照事件的记忆。针对各种潜在的闪光灯记忆事件，已经进行了许多类似的研究，但出于显而易见的原因，研究最多的是2001年9月11日世界贸易中心的恐怖袭击。讽刺的是，我写这篇文章的时间正是2021年9月11日。

Methodologically, flashbulb memory is hard to investigate. Retrospective studies, such as that of Brown & Kulik, can’t settle the accuracy issue because there is no record of participants’ original memories to use as a baseline, nor do we have control, “ordinary” memories from the same period to compare for vividness and accuracy to the flashbulb memory. And, of course, scientists can’t cause world-shaking events just to do experiments. What gradually emerged as a ready-to-go research protocol was to have surveys on hand in case a flashbulb event occurred, in order to quickly interview participants (often students in psychology classes) about their memories for both the flashbulb and a control event. A number of such studies have been performed on various potential flashbulb memory events, but for obvious reasons the most studied one is the terror attack on the World Trade Center on September 11, 2001. Ironically, I’m writing this on 9/11/21.

Talarico 和 Rubin (2003) 于 9 月 12 日对 54 名杜克大学学生进行了结构化访谈，询问他们如何得知袭击事件，并以早期的日常事件（例如参加派对）作为对照记忆。之后，他们在袭击发生后 224 天内每隔一段时间再次采访参与者。他们的研究重点是两起事件记忆的鲜明度与准确性之间的关系。他们以 9 月 12 日收集的第一批回忆作为比较基准。研究结果的核心如下图所示（图 6.7）。

Talarico and Rubin (2003) conducted structured interviews with 54 Duke University students on September 12 about how they learned of the attacks, and, for a control memory, an earlier everyday event, such as attending a party. They then re-interviewed the participants at intervals up to 224 days after the attacks. The focus of their study was the relation between the vividness of the memories of the two events and their accuracy. They used as a baseline for comparison the first recalls collected on September 12. The crux of their results is shown in these graphs (Figure 6.7).

以首次回忆作为衡量记忆准确性的最佳指标，我们发现闪光灯记忆和日常记忆的一致性都会随着时间的推移而降低。也就是说，在Talarico和Rubin所采样的任何时间点，它们的准确性都没有差异。另一方面，闪光灯记忆依然鲜活，并伴有意象，而日常记忆则变得苍白，成为知识而非自传体经验的一部分。简而言之，闪光灯记忆的独特之处在于其鲜活性，而非其准确性；它们之所以看似准确，只是因为它们鲜活（Talarico和Rubin，2007；Hirst和Phelps，2016）。长期的跟踪研究支持了他们的结论（Conway、Skitka、Hemmerich和Kershaw，2009；Curci和Luminet，2006；Hirst、Phelps、Buckner等人，2009；Hirst、Phelps、Meksin等人，2015）。

Taking the first recall as our best measure of memory accuracy, we see that both flashbulb and everyday memories become less consistent over time. That is, there is no difference in their accuracy at any point sampled by Talarico and Rubin. On the other hand, flashbulb memories remain vivid and accompanied by imagery, while everyday memories become wan and part of knowledge rather than autobiographical experience. In short, flashbulb memories are unique in their vividness rather than their accuracy; they seem accurate only because they are vivid (Talarico & Rubin, 2007; Hirst & Phelps, 2016). Long-term follow-up studies have supported their conclusions (Conway, Skitka, Hemmerich, & Kershaw, 2009; Curci & Luminet, 2006; Hirst, Phelps, Buckner et al., 2009; Hirst, Phelps, Meksin et al., 2015).

Lanciano、Curci 和 Semin（2010）就为何生动性不能预测准确性提出了答案。已证实情绪记忆比非情绪记忆更生动，但生动性是一个模棱两可的概念（Brewer，2006）。例如，它是指整体亮度，还是像电视机上设置的特殊色彩平衡？Luciano、Curci 和 Semin 进行了类似闪光灯记忆的实验，发现情绪记忆比非情绪记忆更精细，因为几乎所有体验细节，无论多么无关紧要，最初都会被编码。然而，随着时间的推移，闪光灯记忆会像普通记忆一样发生变化，同时保持同样高水平的细节，因此看起来像照片一样生动。总而言之，与日常记忆不同，它们在主观上保持生动，但与日常记忆一样，它们的准确性会越来越低。

Lanciano, Curci, and Semin (2010) have proposed an answer to why vividness does not predict accuracy. It has been established that emotional memories are more vivid than non-emotional memories, but vividness is an equivocal concept (Brewer, 2006). Is it, for example, overall brightness or a special color balance, as one might set on a TV? Luciano, Curci, and Semin performed experimental analogs to flashbulb memories and found that emotional memories are more fine-grained than non-emotional memories because almost every detail of experience, no matter how irrelevant, is initially encoded. However, over time, flashbulb memories change in the same way as ordinary memories while retaining the same high level of detail and thus seeming photograph-like and vivid. In sum, unlike everyday memories, they stay subjectively vivid, but, like everyday memories, they become less and less accurate.

我们现在的情况

Where We Are Now

我们已经追溯了信息从感觉到记忆的处理过程。尽管关于动物是否具有自传体记忆仍存在一些争议，但我们已经触及了亚里士多德的动物心智和笛卡尔的野兽机器与人类之间的分界线。下一章将探讨传统上被称为“理性”的高级心理过程及其结果——决策。

We have followed the processing of information from sensation through memory. Although there is some dispute about whether animals have autobiographical memory, we have reached the dividing line between Aristotle’s animal mind and Descartes’s beast-machine, on the one hand, and human beings on the other. The next chapter takes up what has traditionally been called the higher mental processes of reason and its outcome, decision making.

外卖

Takeaway

本章的要点是不要相信你的记忆，但更准确地说，认知科学所证明的是记忆的可塑性，而不是遗忘性。奥古斯丁用洞穴来比喻记忆，承认事物可能会在记忆中丢失，存放在洞穴中的瓮可能会破碎并溶解。但他肯定没有想到瓮会变成钟乳石或冬眠的青蛙。从他那个时代到现在，关于记忆的比喻都将其描绘成一个被动的储存库（Draaisma，2000）。然而，对意义的追求并不止于编码。记忆会朝着图式的默认值漂移；每一次记忆活动都会创造一次对意义的新的追求和新的记忆；神经系统将记忆传播到整个大脑，神经过程总是在重新连接神经元之间的连接。记忆是动态的，而不是静态的。从进化的角度来看，记忆的进化是为了满足我们当前的需求，应对适应性挑战，而不是为了保留过去的真实形象。

The takeaway from this chapter is don’t trust your memory, but, more precisely, what cognitive science demonstrates is the malleability of memory, not forgetfulness. Augustine’s caves metaphor for memory recognized that things can be lost in memory, that an urn stored in a cave might break up and dissolve away. But he surely didn’t think that the urn might morph into a stalactite or a hibernating frog. Metaphors for memory from his time to the present have pictured it as a passive repository (Draaisma, 2000). The effort after meaning, however, does not end at encoding. Memories drift in the direction of schemas’ default values; every act of memory creates a new effort after meaning and a new memory; the nervous system spreads memories out over the whole brain, and neural processes are always rewiring the connections between neurons. Memory is dynamic, not static. In evolutionary terms, memory evolved to serve our current needs, to meet adaptive challenges, not to preserve a True image of the past.

阅读材料

Readings

记忆研究的奠基人是赫尔曼·艾宾浩斯（1885/1913）。《记忆：对实验心理学的贡献》（Memory: A Contribution to Experimental Psychology），由HA·鲁格（HA Ruger）和CE·布森纽斯（CE Bussenius）译。（纽约：教师学院出版社）。网址：https ://books.google.com/books/about/Memory.html ?id=oRSMDF6y3l8C，或访问以下网址阅读原文：https://www.google.com/books/edition/%C3%9Cber_das_Ged%C3%A4chtnis/kfA0AAAAMAAJ ?hl=en&gbpv=0 。

The founding work of the study of memory was Herrmann Ebbinghaus, (1885/1913). Memory: A Contribution to Experimental Psychology, trans. H. A. Ruger & C. E. Bussenius (New York: Teachers College Press). Available at: https://books.google.com/books/about/Memory.html?id=oRSMDF6y3l8C, or, for the original: https://www.google.com/books/edition/%C3%9Cber_das_Ged%C3%A4chtnis/kfA0AAAAMAAJ?hl=en&gbpv=0.

弗雷德里克·巴特利特爵士在其著作《记忆：实验与社会心理学研究》（英国剑桥：剑桥大学出版社，1932年）中开创了现代建构主义记忆观。可访问以下网址： https ://pure.mpg.de/rest/items/item_2273030_5/component/file_2309291/content 。

Sir Frederick Bartlett began the modern constructivist view of memory with his Remembering: A Study in Experimental and Social Psychology (Cambridge, UK: Cambridge University Press, 1932). Available at: https://pure.mpg.de/rest/items/item_2273030_5/component/file_2309291/content.

柏拉图曾谴责文字的发明者摧毁了人类的记忆；而目前的嫌疑人是谷歌（Sparrow, Liu, & Wegner, 2011 [讽刺的是，以下链接与上述链接相似。]）。文字出现之前的人们发展出了许多记忆技巧。希腊吟游诗人通过背诵《伊利亚特》和《奥德赛》来娱乐大众。弗朗西丝·耶茨在《记忆的艺术》（纽约：兰登书屋，2011）一书中探讨了文艺复兴时期记忆术的历史。

Plato once denounced the inventor of writing for destroying human memory; the current suspect is Google (Sparrow, Liu, & Wegner, 2011 [ironic following links above.]). Preliterate peoples developed many techniques for remembering. Greek bards provided entertainment with their recitations of the Iliad and Odyssey, learned by heart. Frances Yates discusses the history of mnemonics through the Renaissance in The Art of Memory (New York: Random House, 2011).

学生们经常向专家寻求关于如何提高记忆力的建议。我所知道的唯一一本基于真实研究的书籍是JB Worthen和RR Hunt合著的《记忆术：21世纪的记忆术》（纽约：心理学出版社，2011年）。

Students often ask for expert advice on improving memory. The only book I know based on real research is J. B. Worthen and R. R. Hunt, Mnemonology: Mnemonics for the 21st Century (New York: Psychology Press, 2011).

对目击证人证词可错性的研究，使得认知心理学家被传唤为专家证人（我曾与辩护律师交谈，但从未被传唤）。目击证人证词研究领域最著名的先驱人物是伊丽莎白·洛夫特斯（Elizabeth Loftus），她撰写了一本关于她作为专家证人经历的、类似真实犯罪与心理学结合的著作，名为《辩方证人：被告、目击证人和审判记忆的专家》（纽约：圣马丁出版社，2015年）。

Research on the fallibility of eyewitness testimony led cognitive psychologists to be called as expert witnesses (I once talked to defense lawyers but was never called). The best known, and pioneer of, eyewitness testimony research was (and is) Elizabeth Loftus, who wrote a sort of true-crime-meets-psychologist book about her experiences as an expert witness, K. Ketcham and E. Loftus, Witness for the Defense: The Accused, the Eyewitness, and the Expert Who Puts Memory on Trial (New York: St. Martin’s, 2015).

魔法

Magic

最具争议的魔术分支是心灵魔术。它的兴起源于一场名为唯心论的宗教运动，该运动与早期认知心理学相交叉（Leahey & Leahey，1983）。唯心论始于 1848 年，当时两个无聊的少女福克斯姐妹表演了几个把戏，让很多人相信死者的灵魂依然活着，而且会在房屋里作祟。她们成为了第一批灵媒，唯心论创造了鬼魂的刻板印象和鬼屋的比喻，这些在今天的恐怖电影中仍然存在。唯心论是科学时代的一种宗教，因为它提供了人类个性存在及其在肉体死亡后继续存在的明显证据（Myers，1901 ⁸），而不是将其视为一种信仰。因此，它与心理学、心灵研究、灵魂联系在一起，并成立了专业组织来对后来被称为超自然的精神现象进行科学研究。弗洛伊德认为这是一种职业选择；冯特写了一篇关于德国通灵者的文章；詹姆斯是美国心灵研究协会的创始人之一。统计学首次应用于心理数据的研究是由超心理学家完成的。

The most controversial department of magic is mentalism. It arose because of a religious movement, spiritualism, that intersected with early cognitive psychology (Leahey & Leahey, 1983). Spiritualism began in 1848 with the tricks performed by two bored teenage girls, the Fox sisters, that convinced many that the spirits of the dead live on and can haunt houses. They became the first spirit mediums, and spiritualism created the stereotypes of ghosts and the tropes of haunted houses that endure in horror movies today. Spiritualism was a religion for a scientific age because it offered apparent evidence of the existence of the human personality and its survival after bodily death (Myers, 1901⁸), rather than taking it as a matter of faith. It thus became associated with psychology, the study of psyche, the soul, and professional organizations were formed to carry out scientific investigations of spiritual, later dubbed paranormal, phenomena. Freud considered it a career option; Wundt wrote an article about a German psychic; and James was a founder of the American Society for Psychical Research. The first applications of statistics to psychological data were done by parapsychologists.

不可避免地，一些声称能够帮助幸存者与亲人取得联系的人形成了市场。这些后来被称为“灵媒”的人，有些是真诚的；有些是真诚的，但却用诡计“助长”他们的效果；还有些纯粹是骗子。胡迪尼在招魂术盛行的时期失去了母亲，他寻求灵媒的安慰，但他识破了他们的诡计，并对他们进行了猛烈的攻击。阿尔弗雷德·比奈调查魔术师以击败灵媒（LaChapelle，2008）。也出现了一些魔术师，他们充当娱乐媒介，但声称自己没有超自然力量——心灵感应者。心灵感应的问题之一是，人们常常相信心灵感应者是真正的灵媒，即使他们并非如此。胡迪尼拜访了《福尔摩斯》的作者亚瑟·柯南·道尔，后者（和他的妻子）相信胡迪尼确实拥有通灵天赋，无论胡迪尼如何辩解和演示。福尔摩斯的作者，文学作品中最理性的人物，竟然会相信通灵术（以及其他东西），这证明了人类需要相信。

Inevitably there arose a market for people who claimed to be able to put survivors in touch with their loved ones. Some of the mediums, as they came to be called, were sincere; some were sincere but “helped” their effects with trickery; and some were simply frauds. Houdini, who lost his mother in the heyday of spiritualism, sought them out for comfort, but he recognized their trickery and campaigned ferociously against them. Alfred Binet investigated magicians to defeat mediums (LaChapelle, 2008). There also came into existence magicians who performed as mediums for entertainment but claimed no paranormal powers—mentalists. One of the problems with mentalism is that people often believe mentalists are real mediums even when they are not. Houdini visited Arthur Conan Doyle, the author of Sherlock Holmes, who believed (along with his wife) that Houdini had real psychic gifts, no matter Houdini’s protestations and demonstrations. That the author of Holmes, literature’s most rational being, could fall for psychics (and other things) is a testament to the human need to believe.

心灵魔术在魔术界备受争议的一个主要原因是，它太容易触动人们最脆弱、最敏感的部位，似乎能给予他们无法给予的东西——与挚爱的逝者建立联系。9我在这里提出这^一点因为唯灵论创造了我们随身携带的关于灵魂、闹鬼和灵异现象的图式，而这些图式被一种名为“冷读”（Hyman，2007；Shermer，2010）的虚假灵媒/心灵术所利用。冷读可以让人们认为你比你更了解他们，并引导他们误以为你与他们已故的亲朋好友保持联系，而实际上除了共同的社交图式之外，他们一无所知。10^我个人认为这种魔法是在玩火，我也理解，如果人们被通灵者/心灵术士愚弄了，他们有时会不喜欢魔术师。

A major reason mentalism is controversial among magicians is that it’s too easy to touch people in their tenderest, most vulnerable places, and to seem to offer them something they cannot give, connection to their beloved dead.⁹ I bring this up here because spiritualism created the schemas we carry around about souls and hauntings and spirit manifestations, and these are exploited by a fake medium/mentalist practice called cold reading (Hyman, 2007; Shermer, 2010). It’s possible to make people think you know them better than you do and lead them down a garden path to thinking you are in touch with their dead relatives and friends, without knowing anything at all except for shared social schemas.¹⁰ I personally think this kind of magic is playing with fire and understand that people sometimes don’t like magicians if they’ve been fooled by psychics/mentalists.

最简单的语言就能暗示出图式和情绪。不妨想想史蒂芬·平克提出的以下对话：

Schemas and emotions can be cued by the simplest language. Consider the following dialogs suggested by Steven Pinker:

她：我现在就走。

他：玩得开心。

对比

她：我要离开你了。

他：他是谁？！！

She: I’m leaving now.

He: Have a good time.

vs.

She: I’m leaving you.

He: Who is he?!!

一个词；两种不同的场景/模式。

One word; two different scenarios/schemas.

智力形式

Forms of Intelligence

西蒙（1992，2000）区分了两种智力形式：问题解决和直觉。尽管他受过经济学方面的训练（1981年诺贝尔奖；莱希，1993，2003），但他后来却对经济学感到失望，转而投身心理学。他和艾伦·纽厄尔着手构建一个通用的问题求解器，能够解决从逻辑证明到国际象棋的任何问题（纽厄尔、肖，& Simon, 1959; Ernst & Newell, 1969)。在此模型的基础上，他们构建了人类解决问题的计算机模拟（Newell、Shaw & Simon, 1959; Newell & Simon, 1972）。

Simon (1992, 2000) distinguished two forms of intelligence, problem solving and intuition. Although trained as an economist (1981 Nobel Prize; Leahey, 1993, 2003), he became disenchanted and left it for psychology. He and Allan Newell set out to build a general problem solver capable of solving any problem from logic proofs to chess (Newell, Shaw, & Simon, 1959; Ernst & Newell, 1969). Building on that model, they constructed a computer simulation of human problem solving (Newell, Shaw, & Simon, 1959; Newell & Simon, 1972).

教育家 Carl Bereiter 曾说过，当你不知道该做什么时，你所需要的就是智力。另一方面，直觉依赖于知识，在人工智能中被称为领域特定智能。由于直觉——“跟随你的直觉”——似乎根本不涉及可辨别的心理处理，因此似乎不可能用计算的方式处理它。然而，在研究模式识别时，我们了解到，直观地看到问题的解决方案，或者至少是它们敏锐的表述，取决于深厚的知识，就像专家和新手棋手对棋局的看法之间的差异一样。Kiesel 等人 (2009) 向新手和大师展示了棋盘位置的图片，询问每个棋盘是否有可用的将军步骤。在参与者不知情的情况下，在显示每张图片之前，会显示另一个棋盘位置（素数）并进行遮蔽，以便它永远不会被看到；每个被遮蔽的位置都有或没有可用的将军步骤。研究人员测量了对目标图片做出判断的延迟。新手棋手没有表现出启动条件的影响。然而，当启动条件与目标条件匹配时（即启动条件和目标条件都有或都没有可用的将棋），大师的判断速度更快。专家棋手会对他们看到的局面表现出直觉，但这种直觉是由他们自己没有意识到的认知过程所驱动的。

The educator Carl Bereiter once remarked that intelligence is what you need when you don’t know what to do. Intuition, on the other hand, depends on knowledge, and in AI is called domain-specific intelligence. Because intuition—“going with your gut”—seems to involve no discernable mental processing at all, it might seem impossible to treat computationally. However, in studying pattern recognition we learned that intuitively seeing solutions to problems, or at least their perspicacious representation, depends on profound knowledge, as in the difference between expert and novice chess players’ perceptions of chess positions. Kiesel et al. (2009) showed novices and masters pictures of board positions, asking if each board had or did not have a checking move available. Unknown to participants, before each picture was displayed, another board position was shown (a prime) and masked, so that it was never seen; each masked position had or did not have a checking move available. The researchers measured latency to making the judgment about the target picture. Novice players showed no effect of the primes. However, masters’ judgments were faster when the prime matched the target (i.e., prime and target both had or did not have a checking move available). The experts exhibited intuition about the positions they saw, but this intuition was informed by cognitive processes of which they were unaware.

问题解决：表示和搜索

Problem Solving: Representation and Search

搜索状态空间

Searching the State Space

为了说明 GPS 方法，我们来看一个简单的单人游戏——汉诺塔（Anderson，1993）。你可以买一个，它们有更多塔和/或圆盘的版本（图 7.1）。

To illustrate the GPS approach, we’ll look at a simple one-person game, the Towers of Hanoi (Anderson, 1993). You can buy one, and they come in versions with more towers and/or disks (Figure 7.1).

游戏的目标是将所有三个垫圈从左塔移到右塔，从而重现相同的金字塔结构。每次只能移动一个盘子，并且不能将较大的盘子放在较小的盘子上。与国际象棋不同，游戏中只有一个操作员，负责将盘子移动到塔上。因此，可以显示游戏中所有可能的合法位置。图 7.2显示了汉诺塔问题的完整状态（或搜索）空间。下一步是搜索该空间以找到解（最右边的位置，底行）。

The goal of the game is to move all three washers from the left tower to the right tower, reproducing the same pyramid configuration. You may move only one disk at a time, and you may not place a larger disk on a smaller disk. Unlike chess, there’s only one operator, moving a disk to a tower. It is therefore possible to display all the possible legal positions of the game. Figure 7.2 shows the complete state (or search) space for the Towers of Hanoi problem. The next step is to search the space to find the solution (right-most position, bottom row).

搜索可以是盲目的，也可以是智能的。在盲目搜索中，系统搜索状态空间，而不评估每个新生成的位置对目标移动的帮助程度。在智能搜索中，使用一个或多个评估函数来评估每一步的进展。无论哪种情况，搜索都可以是广度优先或深度优先的。在广度优先搜索中，系统会从整个状态空间中的给定位置生成所有可能的移动。因此，在我们的汉诺塔示例中，系统会生成起始位置，然后生成第二层位置可用的单个走法，依此类推。状态空间会慢慢显现出来。在深度优先搜索中，系统从起始点生成单个走法，然后从起始点生成到第三层位置的单个走法，依此类推，直到到达最后一层。

Search can be blind or intelligent. In blind search the system searches the state space without evaluating how well each newly generated position helps move toward the goal. In intelligent search one or more evaluation functions are used to assess the progress of each move. In either case, search can be breadth-first or depth-first. In breadth-first search, the system generates every possible move from a given position across the whole state space. So, in our Towers of Hanoi example, the two moves available from the starting position would be generated, then the single moves available at the second-level positions would be generated, and so on. The state space would then be slowly revealed. In depth-first search, the system generates a single move from the starting point, then from it generates a single position to the third level and so on until the last level is reached.

广度优先搜索比较保守，因为最终会生成整个状态空间并揭示通往目标的路径。深度优先搜索可能显得“有见地”。在汉诺塔游戏中，系统可能会随机选择从起点到目标的直线位置，从而超越广度优先策略。另一方面，如果它做出错误的选择，就不得不浪费时间回溯到起始位置。最后，搜索可以是正向链接或反向链接。到目前为止的示例都是正向链接，因为我们从起始位置开始并寻找目标位置。在反向链接中，系统从目标位置开始并寻找通往起始位置的路径。反向链接在诸如汉诺塔之类的问题中最有用，因为这些问题具有唯一的解决方案。在国际象棋之类的问题中，存在许多可能的将死，从其中一个将死反向追溯到初始棋盘设置没有多大意义。

Breadth-first searches are conservative because eventually the whole state space is generated and a path to the goal is revealed. Depth-first search can appear “insightful.” In the Towers of Hanoi, the system might at random choose the straight line of positions from start to goal and outrun a breadth-first strategy. On the other hand, if it makes poor choices it will have to waste time backtracking to the starting position. Finally, search can be forward or backward chaining. The examples so far are of forward chaining because we begin at the starting position and look for the goal position. In backward chaining, the system starts with the goal position and seeks a path to the starting position. Backward chaining is most useful in problems such as the Towers of Hanoi, which have unique solutions. In problems such as chess, where there are many, many checkmates possible, working backward from one of them to the initial board setup makes little sense.

智能搜索或剪枝树会尝试避免搜索到达目标的无用或不必要的长路径。汉诺塔（至少在其三塔三盘形式中）非常简单，因此生成整个搜索空间也很容易。然而，在大多数游戏中，组合爆炸会对其产生不利影响。为了应对大多数问题的状态过度增长，尝试找到关闭某些发展路线的方法很有用，这样就不必生成和搜索它们。这样做需要评估移动的比较价值的方法。例如，国际象棋选手已经开发出几种评估位置的指标。其中最简单的指标之一是物质平衡。每个棋子都被分配一个点值：兵值一分，马和象值三分，车值五分，后值九分。因此，在游戏的任何时候，都可以看到哪个玩家领先。例如，如果白方吃掉了车和兵，而黑方吃掉了象和马，则物质平衡为均等。在考虑可能导致换子行动的棋步时，棋手可以通过计算所涉及的点数来评估。白方可能认为可以用马吃掉兵，但意识到马会被黑方的其他兵吃掉，导致净损失两点。由于这是一个糟糕的结果，白方可能决定不这样做，从而从搜索空间中剔除所有相关的棋步。

Intelligent search, or pruning the tree, tries to avoid searching futile or unnecessarily long pathways to the goal. The Towers of Hanoi—at least in its three tower, three disk form—is so simple that generating the entire search space is easy. However, in most games the combinatorial explosion militates against it. To cope with the lush growth of states for most problems, it’s useful to try to find ways to close off some lines of development so they don’t have to be generated and searched. Doing so requires ways to evaluate the comparative worth of moves. For example, chess players have developed several metrics for assessing positions. One of the simplest is material balance. Each chess piece is assigned a point value: Pawns are worth one point, knights and bishops are worth three, rooks are worth five, and the queen is worth nine. Therefore, at any point in the game, it’s possible to see which player is ahead on points. If, for example, White has taken a rook and pawn while Black has taken a bishop and a knight, material balance is even. In considering a move that will result in an exchange of pieces a player can evaluate it by counting the points involved. White might see that it’s possible to take a pawn with a knight but realize that the knight would then be taken by one of Black’s other pawns, resulting in a net loss of two points. Because this is a poor outcome, White might decide not to make it, thus pruning all its attached moves from the search space.

直觉与专业知识：生产系统

Intuition and Expertise: Production Systems

对专业知识的研究使认知科学家相信，智力不仅仅是解决问题的能力（例如，Ericsson & Charness，1994）。还需要对世界的知识，尤其是对某些领域的详细了解。然而，一个艰巨的挑战依然存在，因为大多数关于世界的知识是隐性的，或者正如哲学家迈克尔·波兰尼（1966）在他开创性的书《隐性维度》中所说，是默会的：我们知道的多于我们能说的。事实上，这个问题从一开始就浮现在这本书的背景中，因为它是各种认知架构之间的主要分歧之一。具身认知最有说服力的说法之一是，我们与世界的物理互动产生了不能简化为命题的身体知识。神经网络和预测处理架构将隐性知识定位在我们大脑复杂的神经结构和功能中，这些结构和功能阻止它们转化为符号。

Research on expertise convinced cognitive scientists that there was more to intelligence than problem-solving ability (e.g., Ericsson & Charness, 1994). Knowledge of the world, especially detailed knowledge of some domain, was needed as well. However, a daunting challenge remained because most knowledge of the world is implicit, or tacit as philosopher Michael Polanyi (1966) put it in a pathbreaking book, The Tacit Dimension: we know more than we can tell. In fact, this issue has been floating in the background of this book from its beginning, as it’s one of the major disagreements between the various architectures of cognition. One of embodied cognition’s most persuasive claims is that our physical interactions with the world give rise to bodily knowledge that cannot be reduced to propositions. The neural network and predictive processing architectures locate tacit knowledge in the complex neural structures and functions of our brains that resist translations into symbols.

然而，符号系统方法的倡导者渴望用单一的认知理论创建一个统一的认知科学领域（Newell，1990，1992；Anderson，2007）。为此，他们采用并修改了一种称为生成系统的计算机架构（例如，Newell，1973，1980；Lovett & Anderson，2005；Young，2001）。信息处理模型中知识表征的缺点在于，在被检索到工作记忆之前，它是被动的。然而，隐性长期知识并不通过工作记忆运作，因为在工作记忆中，我们实际上可以讲述我们所知道的一切。Chabris 采访的那位大师能够清晰地表达他的一些专业知识——他曾与心理学家合作，并掌握了诸如组块之类的术语——但这很费力，而且显然，他知道的比能说的要多。认知建模者需要一种方法来激活长期知识，使其能够自行引发行为，而无需加载到工作记忆中。

Nevertheless, advocates of the symbol-system approach were eager to create a unified field of cognitive science using a single theory of cognition (Newell, 1990, 1992; Anderson, 2007). To do this, they adopted and modified a computer architecture called the production system (e.g., Newell, 1973, 1980; Lovett & Anderson, 2005; Young, 2001). The shortcoming of knowledge representation in information processing models is that it’s passive until it has been retrieved into working memory. However, tacit long-term knowledge does not operate through working memory because then we could, in fact, tell everything we know. The grandmaster interviewed by Chabris could articulate some of his expertise—he’d been working with psychologists and had picked up terms such as chunking—but it was hard work and, clearly, he knew more than he could say. Cognitive modelers needed a way to make long-term knowledge active, able to cause behavior on its own without being loaded into working memory.

产生式系统为计算系统添加了新的长期记忆形式——规则记忆，具体来说，是用于称为产生式的规则的记忆。产生式是一种“如果……则”形式的计算规则，它指定一个或一组条件，以及在满足该条件时应采取的行动。例如，如果条件 (a)、条件 (b) 或条件 (c) 为真， 则 采取行动X。

A production system adds a new form of long-term memory to computational systems, rule memory, specifically, for rules called productions. A production is a computational rule of if ... then form that specifies a condition or set of conditions and an action to be taken if the condition or conditions obtain, for example, IF condition (a), and condition (b), or condition (c) is true, THEN take action X.

生产系统通常用于编程视频游戏中的非玩家角色（Waveren，2001；Schiel，nd），例如传奇的 FPS、Quake。3^图 7.3显示了 NPC 生产系统的结构以及它如何在计算周期中运行。

Production systems are often used to program non-player characters in video games (Waveren, 2001; Schiel, n.d.), such as the legendary FPS, Quake.³ Figure 7.3 shows the structure of an NPC’s production system and how it operates through a computational cycle.

计算周期从 NPC“感知”它所处环境的情况开始，包括其健康和武器统计数据、与敌人的距离等因素，掩护地点，以及敌人的健康和武器状态。这些信息被加载到工作内存中，在那里它们成为可能与规则内存中的产生式匹配的潜在条件。例如，一个产生式是如果 敌人可见 并且 （我的生命值< 20% 或 他的武器>>我的）， 那么 提出撤退，其中>>表示“比...强大得多”。假设NPC的敌人（玩家）在走廊尽头可见，并且拥有比NPC强大得多的武器。然后条件将与规则匹配，并且操作将被提交给冲突解决算法，该算法将选择一个提议的操作，然后执行该操作。在这种情况下，NPC将撤退到走廊的一个角落并躲在它后面。这个动作将启动一个新的计算周期，因为游戏情况已经改变，必须重新感知。

The computational cycle begins with the NPC “sensing” what it can see in its situation, including factors such as its health and weapon statistics, distance to the enemy, places in which to take cover, and the health and weapon stats of the enemy. These facts are loaded into working memory, where they become potential conditions that might match a production in rule memory. For example, one production is IF enemy is visible AND (my health is < 20% OR his weapon >> mine), THEN propose retreat, where >> means “much stronger than.” Let’s suppose that the NPC’s enemy (the player) is visible down a corridor and has a much stronger weapon than the NPC. Then the conditions will match the rule and the action will be proposed to a conflict resolution algorithm that will choose one proposed action, which will then be carried out. In this case, the NPC will retreat to a corner in the corridor and hide behind it. This movement will set off a new computational cycle because the game situation has changed and must be re-sensed.

在应用人工智能领域，生产系统过去和现在都被用来通过一种名为知识工程的实践来创建专家系统（Buchanan、Davis、Smith & Feigenbaum，2018）。知识工程师——一位人工智能专家——会与某个领域的公认专家（例如电视剧《豪斯医生》中的人物）面谈，试图提取构成该从业者专业知识的生产规则。为了让专家清晰地表达隐性知识，这个过程通常包括展示专家不熟悉但面试官熟悉的案例，并引出专家的意见，并根据需要提出问题。

In applied AI, production systems were and are used to create expert systems via a practice called knowledge engineering (Buchanan, Davis, Smith, & Feigenbaum, 2018). The knowledge engineer—an AI specialist—sits down with a recognized expert in a field—such as Dr. House from the TV show—and tries to extract the production rules comprising the practitioner’s expertise. To get the expert to articulate tacit knowledge, the procedure usually involves showing cases unfamiliar to the expert but known to the interviewer, and elicits opinions, asking questions as needed.

事实上，豪斯医生就是专家系统的绝佳范例，豪斯医生和他的团队在诊断过程中就体现了这一点。他们的白板就是工作记忆，豪斯医生首先列出本周病例的症状。随着他们所学产生式中的条件被触发，团队提出诊断和行动建议。建议的行动有：抽血、活检等。新的事实被添加到白板上，一些诊断被排除，一些诊断被保留，一些诊断被添加。条件匹配，新的行动被采取，增加了戏剧张力。在某个时刻，豪斯医生通常会被一个看似无关的事件⁴推动，从而恢复一个特别隐性的产生式，决定诊断，并在 41 分钟时治愈病人。

House, in fact, is an excellent example of an expert system in action, as House and his team embody one during their diagnostic sessions. Their whiteboard is working memory, and House begins by listing presenting symptoms from this week’s case. The team suggests diagnoses and actions, as conditions in their learned set of productions are triggered. Actions are proposed: blood drawn, biopsies taken, etc. New facts are added to the board, some diagnoses are eliminated, some remain, and some new ones are added. Conditions are matched and new actions taken, increasing dramatic tension. At some point, House typically is nudged by a seemingly irrelevant event⁴ to recover an especially tacit production, decides on a diagnosis, and cures the patient at the 41-minute mark.

访谈结束后，AI专家会将专家的陈述编入专家系统外壳的生成规则内存中，该外壳包含整个系统的其他结构。专家系统的目标是让专业知识随时随地可用，尤其是在没有维柯丁成瘾的情况下！

After the interview, the AI specialist codifies the expert’s statement into the production rule memory of an expert system shell containing the rest of the structures of the whole system. The goal of expert systems is having expertise available everywhere, especially without a Vicodin habit!

深度学习

Deep Learning

尽管科学界对此充满热情，但联结主义人工智能直到 2010 年代中期所谓的深度学习人工神经网络 (ANN) 兴起才取得进展 (Le Cun，Bengio 和 Hinton，2015；Bengio，2016）。正如我们所见，军事的实际需求促成了人工智能的首次繁荣，而实际需求和资金推动了深度学习的诞生，但在这种情况下，它源于私人利益而非军事利益。Facebook、亚马逊和谷歌等科技公司（1985 年 PDP 宣言的作者之一 Geoffrey Hinton 现在是谷歌的杰出研究员）希望尽可能精准地定位他们的好友推荐和热门故事推荐、产品推荐以及搜索结果和广告投放。

Despite enthusiasm in the scientific community, connectionist AI achieved little until the rise of so-called deep learning artificial neural networks (ANNs) in the mid-2010s (Le Cun, Bengio, & Hinton, 2015; Bengio, 2016). As we’ve seen, it was the practical needs of the military that caused the first flourishing of AI, and practical need and funding drove the creation of deep learning, but in this case, it was via private rather than military interest. Tech firms such as Facebook, Amazon, and Google—where Geoffrey Hinton, one of the authors of the 1985 PDP manifesto, is now Distinguished Researcher—want to target their friending and trending story recommendations, their product recommendations, and their search results and advertisement placements as precisely as possible.

科技公司利用各种模式：点赞模式可能为 Facebook 用户带来新朋友；新闻阅读模式可能推荐某篇文章；购物模式可能推荐新产品；网站链接模式可以对 Google 搜索结果进行优先排序。模式识别是神经网络最擅长的领域，数学和计算能力的进步使得构建和调整大型多层网络（深度学习也因此得名）成为可能，这种网络远比 20 世纪 80 年代和 90 年代的网络强大得多。它们能够非常有效地在数十亿个数据点中寻找模式。

Tech firms exploit patterns: patterns of likes that might suggest a possible new friend for a Facebook user, a pattern of news story reading that might suggest a particular article, a pattern of purchases that might suggest a new product, or a pattern of links between websites to prioritize a Google search result. Pattern recognition is what neural networks do best, and advances in mathematics and computing power have made it possible to construct and tune large, multiple layered networks (hence, the name deep learning), far more powerful than the networks of the 1980s and 1990s. They are very effective at finding patterns among billions of data points.

更具心理学意义的是，一个深度学习计算机程序击败了世界排名第三的围棋大师李世石，在一场锦标赛中五局四胜（Sneed，2016）。围棋是一种战略性棋盘游戏，有点像国际象棋，但在更大的棋盘上进行，棋盘上有黑白棋子，其理念是将棋子放在对手棋子的周围并将其吃掉。国际象棋被认为比围棋简单，李世石的失败令人大吃一惊。国际象棋是由标准符号系统程序掌握的，这些程序每秒可以尝试数百万次移动和反击，以找到最佳移动。然而，众所周知，围棋需要玩家在棋盘上出现的马赛克中看到规律，区分强弱模式，而这是符号系统所不具备的技能。

Of more psychological relevance was the announcement that a deep learning computer program had defeated the world’s number three Go master, Lee Sedol, winning four out of five games in a tournament (Sneed, 2016). Go is a strategic board game, a bit like chess, but played on a larger board on which black and white stones are placed, the idea being to place stones that surround those of your opponent and capture them. Chess is considered to be simpler than Go, and Sedol’s defeat came as a huge surprise. Chess was mastered by standard symbol system programs, which can try out millions of moves and countermoves per second to find the best ones. Go, however, famously requires players to see patterns in the emerging mosaics on the board, distinguishing between weak and strong patterns, a skill that symbol systems don’t have.

获胜的围棋程序名为“Deep Mind”（Silver 等人，2016）。Deep Mind 的数学原理非常复杂，但粗略地说，它是一个混合联结主义/符号系统程序。深度学习网络经过训练可以区分强弱模式，而符号系统组件则生成可能的走法。该程序在与人类对手对弈之前，通过数千次自对弈来学习。虽然 Deep Mind——就像之前的蛮力计算机象棋程序一样——并不试图模仿人类；但它的胜利表明，人工智能正日益展现出与人类水平相当的智能，但这并不意味着拥有类人智能。

The winning Go program is called Deep Mind (Silver et al., 2016). The mathematics of Deep Mind is complex, but roughly speaking it is a hybrid connectionist/symbol system program. Deep learning networks are trained to tell weak and strong patterns apart, while the symbol system components generate possible moves. The program learned by playing against itself thousands of times before playing against human opponents. While Deep Mind—like the brute force computer chess programs that went before it—does not try to imitate humans; its victory shows that AI is increasingly showing human-level intelligence, which does not mean human-like intelligence.

深奥的谜团

Deep Mystery

因此，尚不清楚这些成功对认知科学的贡献有多大，因为深度学习在揭示或解释认知的同时，也制造了不少神秘的事物。一些人工智能科学家甚至将机器学习称为“炼金术”，而人工智能也正经历着自身的复制危机（Hutson，2018，2019）。

Therefore, it’s not clear how much these successes are contributing to cognitive science, because Deep Learning creates at least as much mystery about cognition as it reveals or explains. Some AI scientists even call machine learning “alchemy” and AI is going through its own replication crisis (Hutson, 2018, 2019).

ANN 最严重的问题从一开始就存在。由于它们不是编程的，而是学习的，即使是创造它们的科学家也不知道它们内部发生了什么：它们是不透明的。随着 ANN 越来越深，其中的奥秘也越来越深。在应用中，尤其是在医疗诊断等专家系统中，这个问题被称为可​​解释性失败。使用产生式构建的专家系统可以通过检索产生该行为的相关产生式来解释其行为。ANN 做不到这一点，而相信一个类似魔法系统的诊断会加深人类对人工智能的普遍怀疑。欧盟很快将要求专家系统自我解释，美国国防部高级研究计划局 (DARPA) 正在资助这方面的研究。认知科学的讽刺结果是 Voosen (2019) 所说的“人工智能神经科学”。认知神经科学家必须研究大脑隐藏的工作原理，现在人工智能“神经科学家”必须对他们自己的创造物做同样的事情。

The most serious problem with ANNs has been around since the beginning. Because they are not programmed, but learn, even the scientists who create them don’t know what’s going on inside: they are opaque. As the ANNs get deeper and deeper, the mysteries inside get deeper, too. In applications, especially expert systems such as for medical diagnosis, the problem is known as a failure of explainability. An expert system built with productions can explain its behavior by retrieving the relevant productions that produced the behavior. ANNs can’t do that, and trusting, say, diagnoses to what amounts to a magical system deepens human suspicions about AI in general. The European Union will soon require expert systems to explain themselves and DARPA is funding research on the question. The ironic outcome for cognitive science is what Voosen (2019) calls “AI neuroscience.” Cognitive neuroscientists have to investigate the hidden workings of the brain, and now AI “neuroscientists” have to do the same thing for their own creations.

Zador (2019) 认为，目前的 ANN 无法很好地模拟人类思维，因为它们的制造者忽略了人类智力的关键事实，即人类智力是进化而来的。Zador 指出，ANN 最常见的模式学习方法是监督学习，即 ANN 通过尝试标记照片来做出反应，然后被告知照片中的物体是什么，并通过反向传播调整其连接权重。借助现代计算机，ANN 可以在短时间内处理成千上万个这样的配对。然而，儿童通常看不到如此整齐排列的物体和标签，他们一生中看到的物体和标签的数量也不如 ANN 在几分钟内看到的那么多。然而，幼儿却能轻松快速地掌握事物的名称。ANN 无法很好地模拟人类学习，因为它们缺乏一个关键因素：先天性。ANN 处于学习理论中极端的后天培养端，因为每个 ANN 一开始的连接强度都是随机的——它对世界一无所知。但我们已经看到，现存的哺乳动物的大脑并非如此，它们已经知道，例如，光来自上方的单一光源，而这种知识有助于模式识别。

Zador (2019) argues that current ANNs are poor models for human thought because their makers ignore the key fact about human intelligence, namely that it evolved. Zador points out that the most common method of pattern learning in ANNs is supervised learning, in which the ANN responds to a photo by trying to label it, is then told what the object in the photo is, and adjusts its connection weights via backpropagation. With modern computers, an ANN can go through thousands and millions of such pairs in a short period of time. However, children don’t usually see such neatly paired objects and labels, and they can’t see as many of them in their lifetimes as an ANN sees in a few minutes. Yet very young children readily and rapidly acquire names for things. ANNs are poor models of human learning because they lack something critical: innateness. ANNs are at the extreme nurture end of theories of learning because each ANN starts out with randomized connection strengths—it knows nothing at all about the world. But we have seen that this is not true of existing mammalian brains, which come already knowing, for example, that light comes from above from a single light source, and this knowledge facilitates pattern recognition.

类似的担忧也出现在自然语言处理（即通过图灵测试）的人工智能领域。目前最主流的自然语言处理（NLP）程序是谷歌赞助的 GPT-3。5 ^NLP对于谷歌实现将搜索结果从任何语言翻译成任何其他语言的目标至关重要。GPT-3 可以通过部分图灵测试（Lacker，2020；Cowen，2020b），甚至可以回应哲学家对其性能的批评（Weinberg，2020）。然而，GPT-3 并不使用语法来理解它所读的内容，而是使用一种人工智能版本的古老语言学习方法，称为英语近似法（Deese，1961）。句子中任何单词与可能跟在其后的另一个单词之间存在统计依赖关系。因此，在英语中，“The”后面不会跟着另一个“the”，⁶也不太可能跟着另一个“the”。动词后跟名词，并且后面很可能跟着名词。此外，提示中的单词越多，下一个单词的限制就越大，因此“当技术人员修理调制解调器时，她发现她必须更换……”后面不会跟着“海象”或“眯眯眼”。GPT-3 是一种人工神经网络 (ANN)，它会从互联网上搜索较长的句子，并记录字符串与下一个出现的单词之间的统计依赖关系。当提示一个新的字符串需要翻译时，它会使用统计数据来猜测下一个单词。它给出了可用的结果，但它不是人类语言理解的模型，而且与人类相比，它在接触到错误信息后很难学会说真话^。7

Similar concerns arise in the AI field of natural language processing, i.e., passing the Turing test. The reigning NLP program is Google-sponsored GPT-3.⁵ NLP is important to Google’s goal of being able to translate search results from any language to any other. GPT-3 can pass some Turing tests (Lacker, 2020; Cowen, 2020b), and it can even reply to philosophers’ critiques of its performance (Weinberg, 2020). However, GPT-3 does not use syntax to understand what it reads but uses an AI version of an old verbal learning method called approximations to English (Deese, 1961). There are statistical dependencies between any word and another word that might follow it in a sentence. Thus, the prompt “The” will not be followed by another “the” in English,⁶ is unlikely to be followed by a verb, and is most likely to be followed by a noun. Furthermore, the more words in a prompt the more constrained the next word will be, so “As the technician worked to repair the modem, she found that she had to replace the …” won’t be followed by “walrus” or “squinch.” GPT-3 is an ANN that scrapes the Internet for longish sentences and catalogs the statistical dependencies between word strings and the next occurring word. When given a prompt of a new string to translate, it uses its statistics to guess the next word. It gives usable results, but it’s not a model for human language understanding, and, compared to humans, it has a hard time learning to tell the truth after being exposed to misinformation.⁷

Dunietz (2020) 认为，GPT-3 及其竞争对手代表着人工智能的死胡同。通过图灵测试的竞争已经变成了一场比拼，看谁的 ANN 规模最大，谁能处理最长的字符串。相反，人工智能应该尝试模拟人类的语言理解，Dunietz 建议将故事理解作为工程的试验台，从空间、时间、因果关系和动机等主题来分析故事。这看起来像是回归巴特利特叙事图式的传统，但并未被承认。

Dunietz (2020) argues that GPT-3 and its rivals represent a dead end for AI. The quest to pass the Turing test has become a competition to see whose ANN is the biggest and who can process the longest strings. Instead, AI should try to emulate human language understanding, Dunietz suggesting story understanding as a testbed for engineering, analyzing stories in terms of themes such as space, time, causality, and motivation. This looks like an unacknowledged return to the Bartlett tradition of narrative schemas.

理性的挑战

The Challenges of Rationality

智力长期以来被认为是一种单一的心理能力。在心理测量学中，它是g因素；在非正式信息处理中，它是工作记忆效率。这也是西蒙和纽厄尔的通用问题求解器背后的理念，该程序提出，是可以应用于任何问题（无论其领域）的算法和启发式方法。但我们看到了他们如何意识到专家智能除了依赖于解决问题的技能外，还依赖于特定领域的知识，从而推动了生产系统和知识工程的发展。

Intelligence has long been considered a single psychological capacity. In psychometrics it’s the g factor, and in informal information processing it’s working memory efficiency. It’s the idea behind Simon and Newell’s General Problem Solver, which proposed that there are algorithms and heuristics that can be applied to any problem regardless of its domain. But we saw how they came to realize that expert intelligence depended on domain-specific knowledge in addition to problem-solving skills, leading to the development of production systems and knowledge engineering.

人类思维的通用观点中蕴含着一种假设，即思维是对逻辑和理性的正确运用。逻辑规则在范围和应用上具有普遍性和普遍性；正如斯多葛学派所观察到的，我们可以对任何事物进行逻辑思考。它们是真正的算法：如果正确地应用于真实的前提，它们就能保证得出有效的结论。最后，由于它们体现了理性，因此它们具有规范性，定义了我们应该相信什么。正如 3 + 2 约束我们相信 5 是答案一样，假设p蕴含q且p为真，那么我们必须并且应该相信q也为真。

Nested within the general-purpose view of human thinking is an assumption that thinking is the right use of logic, of reason. The rules of logic are general and universal in scope and application; it’s possible to think logically about anything, as the Stoics observed. They are true algorithms: applied correctly to true premises they guarantee reaching valid conclusions. Finally, because they embody reason, they are normative, defining what we ought to believe. Just as 3 + 2 binds us to believe 5 is the answer, so given that p implies q and that p is true, then we must and should believe that q is also true.

布尔的这段引言表达了认知科学领域这一颇具影响力的观点。布尔并非将他的“符号语言”作为阐明思维的工具，而是将其视为一幅揭示“人类思维的本质和构成”的图景。布尔认为，逻辑不仅为我们提供了思维运作方式的科学描述，也为我们提供了思维运作方式的规范性说明。将这两种观点区分开来，引发了伟大的理性之争。

The quotation from Boole expresses this influential viewpoint in cognitive science. Boole does not offer his “symbolical language” as a tool for clarifying thinking, but as a picture revealing “the nature and constitution of the human mind.” Boole says that logic gives us a scientific description of how the mind does work as well as a normative account of how it should work. Levering these claims apart caused the great rationality debate.

这一过程始于联想主义的发展。休谟认为，实践结论是通过联想而非逻辑得出的。他（1740/1777/1962，第72页）写道：

The process began with the development of associationism. Hume argued that practical conclusions were reached by association, not by logic. He (1740/1777/1962, p. 72) wrote,

由于这种通过心智活动，我们能够从原因推断结果（如巴甫洛夫条件反射），反之亦然，这对于所有人类的生存至关重要，因此，我们不太可能将其信任于……理性，因为理性是缓慢的……在婴儿期根本不会出现任何程度的……并且，在最好的情况下，也只是……极易犯错。

as this operation from the mind, by which we infer effects from causes [as in Pavlovian conditioning], and vice versa, is so essential to the subsistence of all human creatures, it is not probable that it could be trusted to … reason, which is slow … appears not, in any degree, during … infancy, and, at best, is … extremely liable to error and mistake.

“自然智慧”赋予我们这种“本能或机械倾向”，它“运作时绝对可靠”，在出生时就已存在，在其他缺乏逻辑的动物身上也同样存在。因此，形成普遍结论或习惯的能力建立在联想而非逻辑或理性之上，建立在我们从有限事例中进行概括的倾向之上，以及建立在我们认为原因必然与结果相联系的倾向之上。康德对理性至上论的重申，削弱了休谟的影响力，直到20世纪才有所减弱。

The “wisdom of nature” has implanted in us this “instinct or mechanical tendency” that is “infallible in its operation” and appears at birth and in other animals, which lack logic. Thus, the ability to form general conclusions, or habits, is founded on association, not logic or reason, on our propensity to generalize from limited instances, and on our propensity to feel causes necessarily linked to effects. Kant’s reassertion of the primacy of reason blunted the force of Hume’s influence until the 20th century.

向经济学学习

Learning from Economics

一门社会科学探讨了科学（描述“是什么” ）与规范（规定“应该是什么”）之间的区别。经济学家区分了积极的 和规范经济学⁸（例如，弗里德曼，1953）。实证经济学家的工作是发现经济行为规律，就像物理学家的工作是发现物质规律一样。例如，实证经济学家应该能够预测利率变动对经济增长的影响，就像物理学家可以预测施加力对物体的影响一样。你向规范经济学家咨询有关 401(k) 的建议：如果美联储降息 25 个基点，我应该如何调整我的投资组合？显然，这两种经济实践是相互交织的，因为如果没有关于经济行为的预测性科学知识，财务顾问的建议就毫无价值。

One social science has addressed the difference between science, which describes what is, and norms, which prescribe what ought to be. Economists distinguish between positive and normative economics⁸ (e.g., Friedman, 1953). The job of the positive economist is to discover the laws of economic behavior, just as the job of the physicist is to discover the laws governing matter. The positive economist should be able to predict, for example, the effect of an interest rate change on economic growth the way a physicist can predict the effect on an object of the application of force. You consult a normative economist for advice about your 401(k): How should I (ought I) change my portfolio if the Fed drops interest rates 25 basis points? Obviously, these two economic practices are intertwined because, without predictive scientific knowledge about economic behavior, the advice of financial advisers would be worthless.

当我们思考规范时，我们往往首先想到的是那些权威超越自然的道德律令，例如柏拉图的《善的形式》或大多数宗教的造物主上帝。其他规范则由社会自身建立，旨在使事情有序进行，例如制定诸如在道路哪一侧行驶或陪审团需要多少票才能定罪之类的规则。同样，一些规范性规则构成了某种生活方式，通常是一种游戏。例如，在美式橄榄球中，由于担心脑损伤，铲球方式在过去几年发生了很大变化。另一方面，越位规则本身就有助于构成橄榄球这项运动。在美式橄榄球或足球中，你可以越位，但它们不是一回事，在棒球中你也不能越位。

When we think of norms, we tend to think first of moral edicts whose authority lies outside nature, as for example in Plato’s Form of the Good or in most religions’ creator-God. Other norms are set up by societies themselves to get things done in orderly ways, establishing rules such as on which side of the road to drive or how many votes a jury needs to convict. Similarly, some normative rules constitute a certain way of life, often a game. For example, in American football, how tackles may be made has changed a great deal over the past few years because of fear of brain damage. On the other hand, being offside is a rule that by itself helps constitute football as a practice. You can be offside in American football or soccer, but they are not the same thing, and you can’t be offside in baseball.

然而，对于伟大的理性之争来说，至关重要的是一种源于自然本身而非社会或其他外部权威的规范概念。这种规范概念是由自然功能的存在所创造的。例如，一旦确定心脏是一个循环血液的器官（这是它的功能），就有可能探究一颗特定的心脏是否运转良好。正常的心脏能充分地循环血液，压力既不太高也不太低。如果一颗心脏产生的血压过高，可以服用药物来恢复正常功能。如果一颗心脏完全崩溃，可以通过人工植入物来增强其功能，甚至可以更换一颗新的、更正常的心脏。要评估人类是否理性，我们必须首先问，如果有的话，理性服务于什么自然功能。

Critical to the great rationality debate, however, is a concept of norms that derives from nature itself rather than from society or another outside authority. This concept of norm is created by the existence of natural functions. For example, once it was established that the heart is an organ that circulates the blood—that’s its function—it becomes possible to ask whether a given heart is doing a good job or not. A normal heart circulates the blood thoroughly and with neither too much nor too little pressure. If a given heart generates too much blood pressure, medicine can be given to restore normal function. If a heart breaks down completely, it may be augmented by artificial implants or even replaced with a new, more normal, heart. To evaluate whether humans are rational, we must first ask what natural function, if any, reason serves.

一个重要的科学挑战在于，某些器官或系统的功能存在与否，因为有时我们并不清楚这些功能究竟是什么。亚里士多德认为大脑负责冷却血液（它看起来确实像个散热器），而心脏是思考和感觉的中心（我们至今仍会说“我全心全意地爱你”）。在现代神经科学中，早在海马体的记忆功能被阐明之前，它就已经是一个解剖学上可定义的器官了。

An important scientific challenge is posed by the existence of organs or systems that have functions because sometimes it’s unclear what the function is. Aristotle thought that the brain was for cooling the blood (it does look like a radiator) and the heart was the seat of thinking and feeling (we still say, “I love you with all my heart”). In modern neuroscience, the hippocampus was an anatomically definable organ long before its memory functions were worked out.

理性大辩论的参与者们对理性所起的自然功能持有强烈分歧。此外，我们必须考虑理性可能没有自然功能。功能，而是一种在我们大脑进化到一定规模时才出现的能力。一些认知科学家认为意识就是这样的。或者，再举一个例子，我们不太可能为了进行高等物理计算而进化出计算微积分的能力。

Participants in the great rationality debate disagree strongly over what natural function reason performs. Moreover, one must entertain the possibility that it has no natural function but is an ability that just appeared when our brains evolved to a certain size. Some cognitive scientists believe that consciousness is like that. Or, to pose another example, it’s unlikely that the ability to do calculus evolved so we can do advanced physics.

毫不奇怪，认知科学的哲学传统将理性视为我们通向真理的途径。作为认识论者，哲学家们的主要关注点是根除错误，并为人类知识奠定基础。因此，理性的功能在于发现可证成的真信念。布尔在他那个时代代表了这一传统，而福多尔（2000）在我们这个时代则代表了这一传统。在心理学领域，让·皮亚杰（1947、1970、1971）的发生认识论就属于这一传统。皮亚杰的认知发展理论认为，儿童会经历不同的发展阶段，在这些阶段中，心智会构建日益复杂的逻辑形式，从而更好地理解物质世界。当今的心理学家普遍拒绝这种观点，称之为规范主义（Elqayam & Evans, 2011）或知性主义（Mercier, 2016；Mercier & Sperber, 2019）。对皮亚杰的挑战开启了理性之争。

Unsurprisingly, the philosophical tradition in cognitive science viewed reason as our path to truth. As epistemologists, philosophers’ leading concern was rooting out error and establishing the foundations for human knowledge. The function of reason was thus discovering justifiable true beliefs. Boole represented this tradition in his time, and Fodor (2000) does in ours. In psychology, Jean Piaget’s (1947, 1970, 1971) genetic epistemology belonged to it. Piaget’s theory of cognitive development said the child goes through stages of development in which the mind constructs increasingly sophisticated forms of logic that create better knowledge of the physical world. Psychologists today generally reject this viewpoint, calling it normativism (Elqayam & Evans, 2011) or intellectualism (Mercier, 2016; Mercier & Sperber, 2019). Challenging Piaget was the beginning of the rationality debate.

然而，经济学家对理性的功能有不同的看法。对他们来说，人类日常生活的主要目标并非思考真与美（济慈，1820），而是分配稀缺资源，以尽可能地获得幸福。因此，他们关注的不是良好的推理，而是良好的决策，这是理性的标志。通过做出正确的决策，人们能够实现自己的目标，例如成为一名临床心理学家或获得MBA学位。

Economists, however, have a different take on rationality’s function. For them, the primary goal of human beings in everyday life isn’t contemplating Truth and Beauty (Keats, 1820) but allocating scarce resources to be as happy as possible. Hence, they focus not on good reasoning but on good decision making as the hallmark of rationality. By making good decisions people achieve their goals, such as becoming a clinical psychologist or an MBA.

通过关注可衡量的结果，例如收入、财富或社会地位的提升，经济学家为评估理性提供了一种重要的科学工具，即创建量化的理想化模型来预测人们应该如何选择以最大化自身福祉的能力。弗里德曼（1953）在一篇颇具影响力的论文中，让我们设想预测和解释一位台球大师的行为。鉴于台球运动的环境与牛顿力学中物体在无摩擦（或摩擦力极小）的（小）宇宙中运动的环境最为接近，因此，关于选手策略的最佳理论——也就是最能预测选手如何选择和执行击球的理论——是将她的表现归因于对牛顿定律的了解。她不太可能了解这些定律，但她确实体现了这些定律，而我们将牛顿物理学归因于她，比任何其他竞争理论都能更好地解释她的行为。此外，玩家对牛顿原理的了解或体现越好，该玩家在赢得比赛和赚取金钱方面的富裕程度就越高。

By focusing on measurable outcomes, such as income, wealth, or improved social status, economists bring an important scientific tool to bear on evaluating rationality, the ability to create quantitative, idealized models that predict how people ought to choose to maximize their well-offness. In an influential paper, Friedman (1953) asks us to imagine predicting and explaining the behavior of a master billiard player. Given that billiards provides as close to a Newtonian environment of objects moving in a (small) universe of no (or minimal) friction, the best theory of a player’s strategy—that is the one that will best predict how the player chooses and executes her shots—is to attribute her performance to knowledge of Newton’s laws. It’s unlikely that she knows them, but she does embody them, and our attribution to her of Newtonian physics provides a better explanation of her behavior than any rival theory. Moreover, the better a player’s knowledge or embodiment of Newtonian principles is, the better that player’s well-offness will be in terms of games won and money earned.

理性研究中的三个项目

Three Projects in the Study of Rationality

我们将重点关注与重大理性辩论相关的两条研究线索——逻辑推理和决策——围绕斯蒂奇 (2012) 提出的三个理性研究“项目”展开。

We will focus on two lines of investigation bearing on the great rationality debate—logical reasoning and decision making—organized around three “projects” in research on rationality suggested by Stich (2012).

首先是规范性项目：人们应该如何推理和决策？这是哲学家、数学家和规范经济学家的研究课题，旨在阐明推理和决策的正式规范规则。其次是描述性项目。这是心理学家和行为经济学家的研究课题，旨在探究人类的实际推理和决策是否符合规范规则；如果不符合，那么遵循了哪些规则（如果有的话）？第三是政策制定者的补救性项目，旨在提出社会改善或应对人类非理性的方法。

First is the normative project: How ought people reason and make decisions? This is the project of philosophers, mathematicians, and normative economists setting out the formal, normative rules of reason and decision making. Second is the descriptive project. This is the project of psychologists and behavioral economists discovering whether actual human reasoning and decision making conform to the normative rules and, if they do not, what rules, if any, do they follow? Third is the remediation project of policymakers, proposing ways society can improve or cope with human irrationality.

探究推理：Wason 4 卡牌选择任务

Investigating Reasoning: The Wason 4 Card Selection Task

20 世纪 60 年代中期，彼得·沃森（Peter Wason，1966、1968；沃森和约翰逊-莱尔德，1972）提出了一项实验，该实验已被证明是探究人类推理逻辑最稳健的实验（Evans，2017；Ragni、Kola 和约翰逊-莱尔德，2018）。这项实验通常被称为选择任务，并有多种变体。图 7.4展示了它的经典形式。

In the mid-1960s Peter Wason (1966, 1968; Wason & Johnson-Laird, 1972) introduced what has proven to be the most robust experiment for investigating the logic of human reasoning (Evans, 2017; Ragni, Kola, & Johnson-Laird, 2018). Known usually as the selection task, it has been run in many variants. Figure 7.4 shows its classical form.

实验者向参与者展示桌上的这四张卡片，并陈述一条可能适用于这些卡片的规则。参与者被要求选择所有能够证明规则错误的卡片，但不要选择任何不能证明规则错误的卡片。

The participant is shown these four cards on a table, and the experimenter states a rule that may apply to the cards. The participant is asked to choose all cards that can show the rule to be false but none of the cards that cannot.

沃森的灵感源自皮亚杰的认知发展理论。皮亚杰声称，人们在青少年时期会发展出根据命题逻辑原理进行推理的能力。如果真是这样，那么他们应该能够完成这项简单的任务，而这项任务仅仅依赖于理解命题逻辑的一条规则——实质蕴涵，也就是用“如果……那么”语句进行推理。

Wason was motivated by Piaget’s theory of cognitive development. Piaget claimed that during adolescence people develop the ability to reason according to the principles of propositional logic. If so, they should be able to do this simple task that depends on understanding just one rule of propositional logic, material implication, that is, reasoning with “if … then” statements.

蕴涵命题推理的逻辑规范

The Logical Norms of Propositional Reasoning with Implication

在命题逻辑中，两个命题通过一个逻辑运算符连接起来，连接句的真值是一个函数，该函数由真值表定义，该真值表是组成命题的真值模式的函数。图7.5是蕴涵式（也称为条件式）的真值表。用文字表达，该规则为“如果p，则q”。P是蕴涵式的前提；q是蕴涵式的后件。

In propositional logic two propositions are connected by a logical operator, and the truth value of the joint sentence is a function, defined by a truth table, of the pattern of truth values of the component propositions. Figure 7.5 is the truth table for implication, also known as the conditional. In words, the rule is read “If p, then q.” P is the antecedent of the implication; q is the consequent.

请注意，条件断言“如果 p ... 则 q”仅在前提为真（或发生）且后果为假（或未发生）时才为假。因此，在沃森卡片的例子中，唯一可以证伪规则的卡片是正面为 7 而背面为 M 以外的字母的卡片。在使用蕴涵推理时，有两个有效推论和两个无效推论。第一个是肯定前件：给定“如果 p，则 q”为真（即真值表的第 2 行未发挥作用）；被告知 p 为真，因此你可以推断出 q 成立（你必须位于真值表的第 1 行）。第二个有效推论是否定后件：同样，给定“如果 p，则 q”为真（即真值表的第 2 行未发挥作用）；被告知 q 为假，因此你可以推断 p 也为假（你必须位于真值表的第 4 行）。

Note that the conditional claim, “if p … then q,” is false only when the antecedent is true (or occurs) and the consequent is false, or fails to occur. In the case of Wason’s cards, therefore, the only card that can falsify the rule is a card with a 7 on its front and a letter other than M on its back. When reasoning with implication, there are two valid and two invalid inferences. The first is modus ponens: You are given “If p, then q,” as true (that is, line 2 of the truth table is not in play); are told that p is true, and you therefore may infer that q follows (you must be on line 1 of the truth table). The second valid inference is modus tollens: Again, you are given “If p, then q,” as true (that is, line 2 of the truth table is not in play); are told that q is false, so you may infer that p is also false (you must be on line 4 of the truth table).

第一个无效推论是否定前件；你处于肯定前件的情况，但你被告知 p 为假。因为真值表中唯一不起作用的行是第 2 行，所以无论前件的真值状态如何，整体蕴涵都为真。这里有一个例子：在之前的一场足球比赛中，你说如果你心爱的 Fredonia Wombats 队在第四节进攻时争取达阵，那么他们就会赢得比赛。然而，教练选择了轻松的三分球。除非 Wombats 队尝试达阵但输了，否则你不会希望你的主张无效。不尝试达阵使你的主张保持不变，即为真。另一个无效推论是肯定后件。你处于否定后件的情况，但被告知后件 q 为真。但是，真值表中唯一不起作用的行是第 2 行；第 1 行和第 3 行都有效。因此，前件可以是真也可以是假。无法得出任何推论。在足球的例子中，如果有人知道你的预测，并且知道袋熊队赢了（但没看过比赛），他不应该就此断定你的预测是正确的，也就是说，他们在第四次进攻时选择了达阵。他们可能踢进了那个射门得分，然后在之后的6码进攻中获胜。

The first invalid inference is denial of the antecedent; you are in the modus ponens situation, but you are told that p is false. Because the only line of the truth table not in play is line 2, the overall implication is true regardless of the truth status of the antecedent. Here’s an example: Earlier in a football game, you said that if your beloved Fredonia Wombats went for the TD on fourth and goal, then they would win the game. However, coach went for the easy three. You would not want your claim to have been invalidated unless the Wombats tried for the TD and lost. Not trying for the TD left your claim intact, i.e., true. The other invalid inference is affirmation of the consequent. You are in the modus tollens situation, but are told that the consequent, q, is true. Again, however, the only line of the truth table not in play is line 2; lines 1 and 3 are in play. Therefore, the antecedent can be true or false. No inference can be drawn. In the football example, if someone knew your prediction and knew the Wombats had won (but had not seen the game) he should not conclude that your prediction was true, i.e., that they went for the TD on fourth down. They might have kicked the field goal and won later anyway on a pick 6.

现在你应该能够算出 Wason 选择任务的正确答案了。卡片 A 是肯定前件卡，应该选择它，因为如果卡片背面（另一面）出现 X，规则就会被证伪（即你位于真值表的第 2 行）。卡片 D 是否定后件卡，应该选择它，因为如果卡片正面出现 7，你位于真值表的第 2 行，规则必然为假。卡片 B 是否定前件卡，不应选择，因为你可能位于真值表的第 3 行或第 4 行，而规则为真。最后，卡片 C 是肯定后件卡，不应选择，因为你必须位于真值表的第 1 行或第 3 行，而规则为真。

You should now be able to work out the right answer to Wason’s selection task. Card A is the modus ponens card, and it should be chosen because if, say, an X turned up on the back (the other side) the rule would be falsified (i. e., you are on line 2 of the truth table). Card D is the modus tollens card and should be chosen, because if there’s a 7 on the front, you are on line 2 of the truth table, and the rule must be false. Card B is the denial of the antecendent card, not to be chosen, because you could be on either line 3 or 4 of the truth table, and the rule is true. Finally, card C is the affirmation of the consequent card, not to be chosen, because you must be on line 1 or 3 of the truth table, and the rule is true.

简而言之，每个参与者都应该只选择 A 牌和 D 牌作为可能伪造的牌。他们真的这么做了吗？

In short, every participant should pick just cards A and D as the potentially falsifying cards. Do they?

人们是否按照命题逻辑的规范进行推理？

Do People Reason with the Norms of Propositional Logic?

不是。最常见的错误是选择卡片 A 和 C；几乎没有人会选择卡片 A 和 D（参见 Evans，1982 年；Johnson-Laird 和 Byrne，1991 年、1999 年；Wason 和 Johnson-Laird，1972 年）。起初，各种研究者认为选择任务的原始形式由于没有内容（我们现在称之为抽象形式）而过于困难，因此提出了不同的版本。通常，具体版本（我们将这样称呼它们）似乎能带来更好的推理。问为什么成为了伟大理性辩论中的第一个争议。格里格斯的饮酒年龄问题就是一个例子（Cox 和 Griggs，1982 年；Griggs 和 Cox，1982 年、1983 年）（图 7.6）。

No. The most common error is to choose cards A and C; practically no one chooses cards A and D (see Evans, 1982; Johnson-Laird & Byrne, 1991, 1999; Wason & Johnson-Laird, 1972). Initially thinking that the original form of the selection task was too hard because it had no content (we’ll now call it the abstract form), various investigators presented it in alternative versions. Often, the concrete versions (as we will call them) seemed to bring forth better reasoning. Asking why became the first controversy in the great rationality debate. An example is Griggs’ drinking age problem (Cox & Griggs, 1982; Griggs & Cox, 1982, 1983) (Figure 7.6).

早期研究表明，参与者的推理能力会因社会环境而增强，因为社会环境会涉及到发现作弊者——那些通过自私地违反规范而获益的人。因此，在一项实验（Johnson-Laird & Wason，1972）中，扮演邮政检查员的参与者擅长选择违反邮资规定的信封（例如，一等信件的邮票太少）。表面上看，格里格斯的实验似乎也是如此。饮酒者可能只有20岁，而17岁的青少年可能在可乐里加了苏格兰威士忌，这两者都是潜在的违法者。事实上，格里格斯在他位于佛罗里达州的大学发现，参与者会选择符合规范的卡片A和D。然而，当他前往英国重复这项实验时，英国参与者的选择很糟糕。尽管当时英国对饮酒的年龄限制与美国相同，但当时并没有严格执行。9^格里格斯认为，美国参与者只是通过熟悉程度来了解哪些顾客需要被检查：他们是依靠记忆的专家，而不是依靠推理解决问题的能手。然而，他的英国参与者并非美国饮酒规则的专家，他们试图（但很失败地）推理出答案。其他发现也支持格里格斯的观点。系统2承受的认知负荷会导致推理能力下降，而系统1承受的负荷则不会：逻辑推理需要意识（DeWall、Baumeister & Masicampo，2008）。这为“系统2比系统1更不理性”这一令人惊讶的观点提供了重要的初步线索（Oaksford & Hall，2016）。

Earlier research had suggested that participants’ reasoning was enhanced by social contexts in which detecting cheaters—people who benefit by selfishly violating norms—was involved. Thus, in one experiment (Johnson-Laird & Wason, 1972) participants acting as postal inspectors were good at choosing envelopes that violated postage rules (e.g., too little postage on a first-class letter). Superficially, the same seems true in Griggs’ experiment. The drinker with alcohol might be just 20, and the 17-year-old might have Scotch in his Coke, both potential scofflaws. In fact, Griggs found at his university in Florida that participants made the normative choices of cards, A and D. However, when he traveled to Britain and repeated the experiment, British participants chose badly. Although at the time England had the same age restrictions on drinking alcohol as the US, they were not then strictly enforced.⁹ Griggs argued that American participants simply knew by familiarity which patrons needed to be checked: They were experts drawing on memory, not problem solvers drawing on reasoning. His British participants, however, were not experts at American drinking rules, and they tried (badly) to reason their way to their answers. Other findings support Griggs’s idea. Cognitive loads imposed on System 2 cause reasoning to worsen, while loads on System 1 do not: Logical reasoning requires consciousness (DeWall, Baumeister, & Masicampo, 2008). This is an important first clue to the surprising idea that System 2 is more irrational than System 1 (Oaksford & Hall, 2016).

在沃森开创性研究之后的几十年里，选择任务及其乘法变体已在数百名甚至数千名参与者身上进行过测试（参见 Ragni、Kola 和 Johnson-Laird 于 2018 年撰写的综述和荟萃分析）。其中，获得最有力支持的理论是 Johnson-Laird 的模型理论，该理论提出，参与者应构建任务的心理模型，并尝试想象或寻找与既定规则相反的反例。当呈现方式通过熟悉性（如 Griggs 的研究）或明确的道德原则（如果信件是头等信件，则必须贴上头等邮票）或审慎原则（如果实验室技术人员处理血液，则必须戴手套）来辅助这种想象时，推理能力会得到提高。

In the decades following Wason’s pathbreaking work, the selection task and multiplying variants have been run on hundreds, if not thousands, of participants (see review and meta-analysis by Ragni, Kola, and Johnson-Laird, 2018). The best supported theory, Johnson-Laird’s model theory, proposes that participants build a mental model of the task and try to imagine or look for a counterexample to the stated rule. Reasoning improves when the presentation aids such imagining, through familiarity (as in Griggs work) or a clear moral (If a letter is first class it must bear first-class postage) or prudential (If a lab tech is working with blood, he must wear gloves) rule.

无论如何，有一点很明显：知识分子的假设，即人们通常按照逻辑规则从前提推理到结论，是不正确的。

In any event, it is clear that the intellectualist hypothesis, that people routinely reason from premises to conclusions by following the rules of logic is not true.

调查决策：行为经济学

Investigating Decision Making: Behavioral Economics

现在我们来谈谈一个问题，它的解决是塑造现代世界的关键：我们应该如何做出那些结果无法预见的决定？从某种意义上说，我们在本章中，我一直在思考这个问题。棋手并不知道在每种情况下自己都在采取最佳策略，因为对手可能会想到一个她意想不到的策略。然而，在游戏之外更广阔的世界里，决策会带来更重要的后果。想象一下，在18世纪，有人要求你投资从伦敦到中国的商船之旅（Rowe，2010）。你可能会从你的投资中赚到一大笔钱，但如果船在海上沉没，你就会失去一切。那么，理性的，也就是最好的做法是什么呢？

We now come to a question whose resolution was a key to making the modern world: How should we make decisions whose outcome cannot be foreseen? In one sense, we have been considering this throughout the chapter. A chess player doesn’t know that she’s making the best move in each situation because her opponent may think of a strategy that she hasn’t anticipated. However, in the larger world outside games, decisions carry more weighty consequences. Imagine in the 18th century being asked to invest in a merchant voyage from London to China (Rowe, 2010). You might make a fortune on your investment, but if the ship goes down at sea, you lose everything. What’s the rational, i.e., best, thing to do?

决策与博弈的经济规范

The Economic Norms of Decision Making and Games

这个问题最早由17世纪的法国数学家提出。一些数学家受雇于渴望提高赌博胜率的贵族，但其核心人物是布莱斯·帕斯卡（1623-1662）。他是一位数学家（他发明了最早的机械计算器之一）和科学家，但由于宗教信仰危机，他隐居修道院。在那里，他参与撰写了一本影响深远的著作《逻辑，或思考的艺术》（Arnauld & Nicole，1683年）。正如书名所示，这本书以布尔式的语气开篇：“逻辑是正确引导推理以获得事物知识的艺术”（第24页）。帕斯卡的贡献是迈向经济学的第一步，即在不确定条件下进行决策的科学。帕斯卡让无神论者就上帝的存在打赌，他希望说服无神论者相信信仰上帝是理性的（Pascal，1670年）。

This question was first taken up by French mathematicians in the 17th century. Several were employed by aristocrats eager to improve their chances at gambling, but the central figure was Blaise Pascal (1623–1662), mathematician (he built one of the first mechanical calculators) and scientist, but who, due to a crisis of religious faith, retired to a monastery. There, he participated in writing an influential book, Logic, or the Art of Thinking (Arnauld & Nicole, 1683). As its title suggests, the book begins in a Boolean tone: “Logic is the art of directing reasoning aright in attaining the knowledge of things” (p. 24). Pascal’s contribution was the first step toward a science of economics, of decision making under uncertainty. Pascal asked atheists to place a bet on the existence of God and he hoped to persuade atheists that belief in God was rational (Pascal, 1670).

帕斯卡赌注的结构为决策理论提供了模板。你需要在几种不同的行动方案中做出选择，每种方案都会带来不同的后果。在帕斯卡赌注中，有两个选择：相信上帝的存在或拒绝上帝的存在。世界有两种可能的状态：上帝存在或不存在，每种状态都会给你带来不同的结果或收益（图7.7）。

The structure of Pascal’s wager provided the template for decision-making theory. You have a choice to make among several alternative courses of action, each of which has different consequences. In Pascal’s wager, there are two choices: Believe in God’s existence or reject it. There are two possible states of the world: God exists or does not exist, and each has different outcomes, or payoffs, for you (Figure 7.7).

在第一栏中，我们看到，无论你是否相信上帝，上帝的不存在都不会对你人生的幸福产生多大影响——你会享受尘世的生活。然而，上帝可能存在，却会对死后的生活产生巨大的影响。如果上帝存在，你将在天堂体验永恒的幸福。如果上帝不存在，你将在地狱遭受永恒的痛苦。对帕斯卡来说，无神论者的选择应该很明确：相信相信上帝。当帕斯卡发现这场赌注无法使任何人改变信仰时，他是第一个对人类理性感到绝望的人。

In the first column, we see that whether you believe in Him or not, God’s non-existence would make little difference to your life’s happiness—you would enjoy life on earth. However, God’s possible existence makes a big difference after death. If God exists, you will experience an eternity of bliss in heaven. If God does not exist, you will suffer an eternity of pain in hell. To Pascal, the atheist’s choice should be clear: Believe in God. Pascal was the first person to despair of human rationality when he discovered that the wager made no converts.

贵族们的数学专家们通过添加概率、期望和效用的概念，对帕斯卡的简单赌注进行了进一步的阐述。人们至少从古埃及就开始赌博了，在古埃及的墓葬中，我们发现了刻有铭文的骨头，它们是骰子的前身。然而，似乎直到17世纪才有人将概率视为一个可计算的值。在帕斯卡的赌注中加入概率，可以计算出如何在经济学中称之为“不确定性”的条件下做出决策。不过，我将遵循塔勒布（2007）和吉格伦泽（2008）略有不同的用法：在确定性、风险和不确定性条件下进行决策。

The aristocrats’ mathematical boffins elaborated Pascal’s simple wager by adding the concepts of probability, expectation, and utility. People have been gambling at least since ancient Egypt, where among tomb goods we find incised bones that were precursors of dice. Nevertheless, it appears that no one thought of probability as a computable value until the 17th century. Adding probability to Pascal’s wager makes it possible to calculate how to make decisions under uncertainty, as it’s called in economics. However, I am going to follow Taleb’s (2007) and Gigerenzer’s (2008) somewhat different usage: decision making under certainty, risk, and uncertainty.

• 确定性。一切都是可量化且确定的，其结果可以被完美预测。这就是经典物理学的世界。
• Certainty. Everything is quantifiable and deterministic, and outcomes can be perfectly predicted. This is the world of classical physics.
• 风险。结果并不确定，但可以通过概率计算。这是一个（诚实的）机会游戏的世界，比如掷骰子。如果你认为自己在一个确定的世界中运作，而实际上你在一个充满风险的世界中运作，就会产生零风险的错觉。例如，在医学领域，测试结果通常被认为是确定的，但事实并非如此，它们是信号检测理论的实验室版本。每项测试都会漏掉一些患病的人，并错误地警告一些没有患病的人。在艾滋病刚出现的时候，许多接受过一次阳性检测的人被告知他们患有艾滋病，而实际上并没有，这导致了不必要的痛苦和毫无意义的治疗。当公共卫生需要大规模检测时，处理假阳性是一个重要问题，就像 COVID-19 一样。
• Risk. Outcomes are not certain but can be probabilistically calculated. This is the world of (honest) games of chance such as craps. Thinking you are operating in the world of certainty when you are actually operating in a risky world creates the zero-risk illusion. In medicine, for example, test results are often presented as definitive, but they are not, being lab versions of signal detection theory. Every test will miss some people who have a disease and falsely alarm some people who don’t. When AIDS was new, many people who had received a single positive test were told they had AIDS when they didn’t, leading to unnecessary anguish and pointless treatments. Dealing with false positives is an important issue when public health demands mass testing, as with COVID-19.
• 不确定性。结果并非确定，也无法用概率计算。如果以为自己身处风险世界，而实际上却在充满不确定性的环境中，就会产生“火鸡错觉”。一个人每天都来喂一只火鸡，火鸡每次喂食都会欣喜若狂。然而，11月下旬的一天，这位好心人手里却拿着一把斧头。鉴于这场关于理性的伟大辩论，值得注意的是，火鸡对食物的期待完全合理，因为在那个决定性的感恩节之前，这一预测一次又一次地被证实——它会产生强烈的巴甫洛夫条件反射 (CR)。尽管2008年的金融危机更为复杂，但部分原因也是由“火鸡错觉”造成的。投资者购买了将多笔房屋抵押贷款捆绑在一起的金融工具。其理念是，止赎是随机事件，因此即使一些抵押贷款出现问题，规模更大的工具仍能保值。然而，就像火鸡对食物的期望一样，随着一波违约潮的突然出现，投资者的假设被证明是悲剧性的错误（Lewis，2010；另见根据刘易斯改编的电影《大空头》）。
• Uncertainty. Outcomes are not certain and cannot be probabilistically calculated. Thinking you are operating in the world of risk when you are actually operating under uncertainty leads to the turkey illusion. A man comes and feeds a turkey every day, and the turkey comes to happily expect each feeding. However, one day in late November the nice man has an axe in his hand. In light of the great rationality debate, it’s important to note that the turkey’s expectation about food was perfectly justifiable, as the prediction was confirmed over and over again—it would have a strong Pavlovian Conditioned Response (CR)—before that fateful Thanksgiving. Although the financial meltdown of 2008 was more complex, part of it was caused by the turkey illusion. Investors bought financial instruments that bundled many home mortgages together. The idea was that foreclosures were random events, so that if a few mortgages went bad the larger instrument would retain value. However, like the turkey’s expectation of food, investors’ assumption turned out to be tragically wrong as a wave of defaults occurred all at once (Lewis, 2010; see also The Big Short, the movie based on Lewis).

表7.1展示了帕斯卡赌注的一个版本，其中引入了期望值的概念。这是一个小版本的问题：我应该投资朋友的贸易企业吗？

Table 7.1 shows a version of Pascal’s wager introducing the concept of expectation. It’s a small version of the question, should I invest in my friend’s trading venture?

表 7.1风险下的决策

你有一个选择，它有两个不确定但可计算的结果。每个选择的预期值是通过将每一行的收益或损失乘以其概率，然后将所有行的结果相加得出的。理性选择是预期值较高的选择。

You have a choice that has two uncertain—but calculable—outcomes. The expected value of each choice is calculated by multiplying the gain or loss by its probability in each row, then summing the results across the rows. The rational choice is the one with the higher expected value.

最后，必须添加风险规避和效用的概念才能全面了解理性经济人，即理性决策者。想象一下，一天下午我遇到你，你说你饿了。我有一些薯片袋，可以给你一份，你接受了。这并不是一个艰难的决定。但假设我说，我可以给你一个交易：1.25 美元肯定可以让你得到一袋薯片，或者你可以赌 1.25 美元抛硬币，如果正面朝上，你可以得到两袋，但如果反面朝上，你就得不到薯片。如果你通过决策表运行这个，每次下注的预期值都是一袋薯片（p = 1 × 1 袋 = 1 袋预期值 vs. p = 0.5 × 2 袋 = 1 袋预期值）。你会选择哪一个？你应该无所谓，但大多数人会选择肯定的事情——他们厌恶风险。结论是，结果不仅仅与金钱有关，还与什么是有用的、什么对你有效用有关。作为一个饥饿的人，对于你来说，获得薯片肯定比用同样的钱得不到薯片的可能性更有效用。最后，请注意，你当前的财富会改变效用。假设你的背包里有 50 袋薯片，我提出我的报价。你很可能会冒险再买两袋，因为你甚至不需要一袋。这就是边际效用递减的思想。如果你没有薯片，一袋薯片的效用很高，但如果你有很多袋薯片，下一袋薯片的效用就不高（在边际上）。你拥有的东西越多，下一单位薯片的价值就越低。

Finally, the concepts of risk aversion and utility must be added to get the full picture of Homo economicus, man the rational decision maker. Imagine that I come across you one afternoon, and you say you’re hungry. I’ve got some bags of potato chips and offer you one, and you take it. Not a hard decision. But suppose instead I said, I’ll offer you a deal: $1.25 gets you one bag of chips for sure, or you can bet $1.25 on a coin flip such that if it’s heads you get two bags, but if it’s tails you get no chips. If you run this through the decision table, the expected value of each bet is one bag of chips (p = 1 × 1 bag = 1 bag expected value vs. p = 0.5 × 2 bags = 1 bag expected value). Which do you choose? You should be indifferent, but most people choose the sure thing—they are averse to risk. The conclusion is that outcomes are not just about money, they are about what is useful, what has utility for you. As a hungry person, getting chips for certain has more utility to you than the possibility of no chips for the same money. Finally, note that your current wealth changes utilities. Suppose you have 50 bags of chips in your rucksack and I make my offer. You might very well take the risk of getting two more because you don’t need even one very much. This is the idea of diminishing marginal utility. If you have no chips, one bag has high utility, but if you’ve got lots of bags, the next bag (at the margin) doesn’t. The more you have of something the next unit of that thing loses value.

18世纪，杰里米·边沁（Jeremy Bentham，1748-1832）试图将经济学家的行为效用概念转化为心理学概念（Bentham，1789），他声称效用只是快乐的另一种说法，如果是负值，则为痛苦。因此，人们应该能够构建一个心理物理的快乐/痛苦量表，以应对那些不涉及金钱。信奉边沁功利主义的思想家有时会谈到幸福的“享乐”或痛苦的“痛苦”。想象一下，你在一个阴沉的早晨醒来，想知道是否应该带上一把伞。你权衡着带伞的享乐和痛苦，淋湿还是保持干燥，以及可能丢失伞等等，如表7.2所示。

In the 18th century, Jeremy Bentham (1748–1832) tried to turn economists’ behavioral concept of utility into a psychological one (Bentham, 1789), claiming that utility is simply another name for pleasure, or when negative, pain. Thus, one should be able to formulate a psychophysical scale of pleasure/pain to address choices that don’t involve money. Thinkers who follow Bentham’s utilitarianism sometimes speak of “hedons” of happiness or “agons” of pain. Imagine that you wake up one grey morning and wonder if you should pack an umbrella. You weigh the hedons and agons of carrying it, getting wet vs. staying dry, possibly losing it, etc., as shown in Table 7.2.

表 7.2我应该采取 bumbershoot 吗？

您应该选择带来最少痛苦或最大快乐的结果。

You should choose the outcome leading to the least pain or greatest pleasure.

边沁的观点颇具牛顿式的风格，也是他提出的改良方案中的第一个提案。他提出，我们是理性的，我们运用他提出的幸福演算——幸福演算——来做出决策，但如果我们将其量化，就能提升我们的推理能力，就像炮兵通过反复试验学会了以中等精度射击，但通过运用经典物理学，他们的射击精度得到了极大的提升。边沁的最终目标不仅是改善个人决策，更是改善政府的政策制定，而这正是当今理性大辩论的焦点^。10

Bentham’s point of view here is rather Newtonian and is the first proposal in the ameliorative project. He proposed that we are rational, making decisions by means of his felicific calculus—calculus of happiness—but that we can improve our reasoning if we quantify it, just as artillerymen who by trial and error had learned to shoot with modest accuracy greatly improved it by employing classical physics. Bentham’s ultimate goal was not just improving individual decision making but also governmental policymaking, a focus of today’s great rationality debate.¹⁰

这个问题由来已久：决策者应该如何做出对整个社会最有利的决策？功利主义以新的视角看待这个问题。个人天生自私，追求自身幸福，而这又是科学真理，无法通过设定公正的监护人或信赖君主的高瞻远瞩来回避。立法机构并非好到哪里去，因为它们是由一群自私的个体组成的（研究它们构成了经济学的一个分支，称为公共选择理论；Buchanan & Tullock, 1962）。边沁的愿景是运用幸福演算，将所有公民的幸福加总，从而计算出最多人的最大幸福，从而消除人类个体的个体判断。

The problem was ancient: How should policymakers make decisions that are best for their whole society? Utilitarianism placed the problem in a new light. Individuals are naturally selfish, pursuing their own happiness, and, as this is a scientific truth, it can’t be evaded by positing disinterested guardians or trusting to the elevated perspective of monarchs. Legislatures aren’t much better because they are collections of selfish individuals (studying them constitutes a branch of economics called public choice theory; Buchanan & Tullock, 1962). Bentham’s vision was to eliminate idiosyncratic human judgments by using the felicific calculus to sum across the pleasure of all citizens, computing the greatest happiness for the greatest number.

到目前为止，我们讨论的都是一些与自然结果（例如抛硬币和天气）相关的决策。随着计算机的发展，以及在许多情况下引入了相同的思想家（冯·诺依曼和摩根斯特恩，1944），决策理论也得到了扩展，扩展到你的选择会影响他人选择的情形，例如国际象棋或扑克等游戏。如果再增加一个人，决策理论将进入一个更加复杂、更具心理维度的空间，因为你不仅要考虑是否会下雨，或者是否会正面朝上，还要考虑其他玩家会如何回应你的选择。

So far, we have considered decisions that pit you against natural outcomes, such as coin flips and weather. At the same time as computers were being developed and including in many cases the same thinkers (Von Neumann & Morgenstern, 1944), decision-making theory was extended to cases in which the choice you make affects the choices of others, as in games such as chess or poker. Adding another human being takes decision theory into a more complex and psychological space, as you must consider not just whether it will rain or not, or if heads will come up or not, but how the other player will respond to your choices.

博弈论推理最著名的例子是囚徒困境，我们在《法律与秩序》等犯罪剧中经常看到。两个人因犯罪被捕，每个人都面临一个交易：要么牵连他们的同伙，要么冒着同伙背叛的风险。正如杰克·麦考伊曾经说过的：“第一个举手的人就能成交！” 我们可以用一个常用的表格来概括这个决策过程（表 7.3）。

The most famous example of game theory reasoning is the prisoner’s dilemma, which we see in crime shows such as Law and Order. Two people are arrested for a crime, and each is offered a deal to implicate their partner in crime or risk that the partner will betray them. As Jack McCoy used to say, “First one to raise their hand gets the deal!” We can use the usual table to sketch the decision (Table 7.3).

表 7.3囚徒困境

在这种情况下，我们无需计算期望值，即可根据决策理论评估我们的选择。假设你是斯泰西。如果你选择保持沉默，你最好的结果就是入狱一年，前提是（！）约翰也保持沉默。如果你牵连约翰，你的另一个选择，你最好的结果就是获得自由，前提同样是约翰保持沉默。考虑到这个收益矩阵，以及经济学原则——你始终应该做出效用最大化的选择——理性的选择是背叛——牵连约翰。此外，你意识到约翰也面临着同样的交易，对他来说，理性的选择是也包括背叛，所以你应该尽快背叛。然而，请注意，如果我们将斯泰西和约翰视为共同伴侣，视角就会发生变化。最糟糕的共同结果是双方都背叛（四年监禁），其次是一方告发另一方（三年监禁），最好的结果是双方都保持沉默（两年监禁）。纯粹自私的效用最大化意味着背叛，而公共效用最大化则意味着利他主义。多数人的效用应该大于少数人的效用吗？哪种选择是理性的，尚无定论（Ostrom，2009）。

We don’t need to compute expectations in this case to evaluate our choices from decision theory. Imagine you are Stacey. If you choose to stay silent, your best outcome is one year in prison, assuming (!) that John stays silent, too. If you implicate John, your other option, your best outcome is going free, again assuming that John stays silent. Given this payoff matrix, and the economic principle that you should always make the choice that maximizes your utility, the rational choice is defection—implicate John. Moreover, you realize that John is being offered the same deal, for whom the rational choice is also defection, so you should defect as soon as possible. Observe, however, that if we consider Stacy and John as a joint partnership, the perspective changes. The worst joint outcome is for both to defect (four years in prison), the next worst is for one to inform on the other (three years in prison), and the best outcome is for both to stay silent (two years in prison). Purely selfish utility maximization suggests defection, while communal utility maximization suggests altruism. Should the utility of the many outweigh the utility of the few? Which choice is rational is up for grabs (Ostrom, 2009).

偏见和启发法

Biases and Heuristics

启发式方法和偏见的清单很长，而且还在不断增长。12^在这里，我将分别列出其中的几种。

The list of heuristics and biases is long and growing.¹² Here, I will list a few of each.

启发式和偏见是缺陷还是特性？

Are Heuristics and Biases Bugs or Features?

除了得出人们遭受“可补救的非理性思维”的结论外，我们还能如何解释“战争、经济衰退……宗教狂热……[和]破裂的婚姻”？

Stanovich，2004，第159页

How can we explain “wars, economic busts…religious fanaticism…[and] broken marriages” except by concluding that people suffer from “remediable irrational thought?”

Stanovich, 2004, p. 159

改良主义

Meliorism

在会计和经济学中，沉没成本是指无法收回其价值的投资；例如，你在度假游轮上支付的不可退还的订金——无论你去不去，都无法取回。沉没成本谬误是指你让未来的决策受到沉没成本的影响。它在人类和组织中非常常见（Arkes & Blumer，1985），甚至在动物中也非常常见（Sweis et al.，2018）。

In accounting and economics, a sunk cost is an investment whose value cannot be recovered; an example is a non-refundable deposit such as you might make on a vacation cruise—whether you go or not, you can’t get your money back. The sunk cost fallacy occurs when you let a decision about the future be affected by consideration of sunk costs. It is very common in people and organizations (Arkes & Blumer, 1985) and even in animals (Sweis et al., 2018).

例子比比皆是。一个著名的例子给这个谬误起了另一个名字，即协和式飞机谬误。20 世纪 60 年代，英国和法国政府成立了一家公司，制造超音速运输机协和式飞机。但在中途，他们意识到这架客机永远无法盈利，因为研发的沉没成本无法偿还。法国人和英国人并没有放弃，他们说我们花了太多钱，不能放弃，就像你生病了，待在家里会更好，但还是会去坐邮轮一样，他们说如果你不去，你的订金就浪费了。但协和式飞机的成本和你的订金都已经损失了——浪费了——而且无法收回。协和式飞机的投资者为了追逐已经消失的资金而损失了更多的钱，而你在船上过得痛苦不堪，也失去了效用。即使不涉及金钱，也会有沉没成本（Olivola，2018）；恋爱关系持续的时间越长，分手就越难。

Examples abound. A famous one gave the fallacy another name, the Concorde Fallacy. In the 1960s the British and French governments set up a company to build a supersonic transport, the Concorde. It became clear mid-way that the airliner would never be profitable because the sunk costs of development could not be repaid. The French and British didn’t give up, saying we spent too much to quit, just as you might go on your cruise even though you’ve become ill and would be better off at home, saying that if you don’t go, you’ll have wasted your deposit. But the Concorde’s cost and your deposit are already lost—wasted—and can’t be recovered. The Concorde’s investors lost still more money chasing money already gone, and you will lose utility by being miserable aboard ship. Even when money isn’t involved sunk costs occur (Olivola, 2018); the longer a romantic relationship has lasted, the harder it is to break up.

尽管谚语“覆水难收”、“莫把钱扔到九霄云外”，以及莎士比亚名言“木已成舟，既已成舟”都提醒我们要避免沉没成本谬误，但显然我们还需要进一步的补救措施。在这种情况下，研究发现，进行沉没成本方面的短期培训可以减少错误的沉没成本思维（Larrick、Morgan 和 Nisbett，1990；Strough 等，2016）。理查德·尼斯贝特在一次演讲中将理解这一理念描述为“像病毒一样”，可以改变你的整个人生。他说，这番话听起来颇具禅宗（或斯多葛学派）的韵味：“你的整个人生从现在开始！”

Although proverbs, “It’s no use crying over spilt milk,” and “Don’t throw good money after bad,” and a Famous Shakespearian Quote, Lady Macbeth’s “What’s done is done,” remind us to avoid sunk cost fallacies, it’s clear further remediation is needed. In this case, brief training in sunk costs has been found to reduce fallacious sunk-cost thinking (Larrick, Morgan, & Nisbett, 1990; Strough et al., 2016). In a talk, Richard Nisbett described grasping the idea as working “like a virus” that can change your whole life. He said the lesson sounds rather Zen [or Stoic], “Your whole life starts right now!”

助推

Nudging

这是泰勒和桑斯坦（2008）创造的一个术语。其理念遵循斯金纳的理论：创造一种环境，温和地推动人们做出更好的决策，但不要公开透明地表达出来（de Ridder、Kroese 和 van Gestel，2021）。他们最喜欢的例子是退休规划。通常，当人们被聘用从事新工作时，他们会填写一份包含个人数据的电脑表格。其中一个问题是：“你想加入 401(k) 养老金吗？”表格通常默认选项填写为“否”。但他们表示，应该填写“是”。这并不剥夺员工做出其他决定的能力——这不是强制性的家长式作风——而是会引导员工朝着正确的方向发展。这种安静的家长式作风通常比强制性家长制（Sunstein，2017），但有时它不起作用，命令变得更可取，例如戴口罩以防止大流行疾病传播的情况（Sunstein，2020）。

A term coined by Thaler and Sunstein (2008). The idea is Skinnerian: Create environments that gently push people into making better decisions, without being open and obvious about it (de Ridder, Kroese, & van Gestel, 2021). Their favorite example involves planning for retirement. Often when people are hired for new jobs, they fill out a computerized form with their personal data. Among the questions is, “Do you want to sign up for a 401(k) fund?” The form usually has the default choice filled in as “No.” But it should, they say, be filled in with “Yes.” This does not remove the person’s ability to decide otherwise—it’s not coercive paternalism—but does nudge the employee in the right direction. This kind of quiet paternalism is usually preferable to coercive paternalism (Sunstein, 2017), but it sometimes doesn’t work, and commands become preferable, as in the case of wearing masks to prevent the spread of pandemic disease (Sunstein, 2020).

改革机构程序

Reforming Institutional Procedures

这是 Gigerenzer（Gigerenzer 和 Kurzenhäuser，2005；Gigerenzer 和 Gaissmeier，2011；Raab 和 Gigerenzer，2015）特别关注的领域。他认为，通过应用快速简便的启发式方法取代医生不信任的直觉和数学模型（或人工神经网络），可以改善医疗等机构的决策。例如，在心脏医学领域，Gigerenzer 的团队已经证明，在实时决策如何治疗心脏病患者时，简单的三步决策树比直觉或方程式更快、更可靠、更准确。航空公司飞行员会遵循详尽的程序清单，以避免记忆失误并减少认知负荷。

This has been a special focus of Gigerenzer (Gigerenzer & Kurzenhäuser, 2005; Gigerenzer & Gaissmeier, 2011; Raab & Gigerenzer, 2015), who believes decision making in institutions such as medicine may be improved by the application of fast and frugal heuristics to replace physicians’ intuitions and mathematical models (or ANNs) they don’t trust. For example, in cardiac medicine, Gigerenzer’s team has shown that a simple three-step decision tree is much faster, more reliable, and more accurate than intuition or equations in making real-time decisions about how to treat heart patients. Airline pilots follow extensive checklists of procedures to avoid memory lapses and reduce cognitive load.

我们现在的情况

Where We Are Now

人们创造自己的历史，但这并不是随心所欲地创造的；他们不是在自我选择的环境下创造的，而是在已经存在的、既定的、从过去传承下来的环境下创造的。

卡尔·马克思，路易·拿破仑的雾月十八日 (1851) ¹⁴

Men make their own history, but they do not make it as they please; they do not make it under self-selected circumstances, but under circumstances existing already, given and transmitted from the past.

Karl Marx, 18th Brumaire of Louis Napoléon (1851)¹⁴

让我们尝试理解“理性大辩论”，并转向弗农·史密斯（2002，2008）继亚当·斯密（1776）、哈耶克（例如，1948）和西蒙（1981）之后，对建构理性和生态理性所做的区分。古典经济理论家运用科学理性构建逻辑和数学模型，以预测和解释现代条件下的人类思维和决策。然而，史密斯指出，他们随后犯了一个笛卡尔式的错误（哈耶克，1948），认为“社会制度是并且应该由人类理性的自觉演绎过程创造”（第505页）。对建构主义者来说，思考和决策是双重理性的：社会制度是理性建构的，而那些理论描述和解释行为的理论是由社会科学家理性建构的。因此，正如规范主义者所言，人类理性在于遵循逻辑和决策规则。

Let us try to make sense of the Great Rationality Debate by turning to a distinction that Vernon Smith (2002, 2008), following Adam Smith (1776), Hayek (e.g., 1948), and Simon (1981), makes between constructive rationality and ecological rationality. Classical economic theorists used scientific reason to construct logical and mathematical models to predict and explain human thinking and decision making under modern conditions. However, they then made a Cartesian mistake (Hayek, 1948), Smith says, by believing that therefore, “social institutions were and should be created by conscious deductive processes of human reason” (p. 505). To the constructivist, thinking and deciding are doubly rational: Social institutions were rationally constructed, and the theories that describe and explain behavior were rationally constructed by social scientists. Therefore, as the normativists would have it, human rationality consists in following the rules of logic and decision making.

但这种解释本末倒置，把笛卡尔置于马之前。马就是历史。斯密从马克思著名的洞见入手（尽管斯密将其归功于哈耶克，1948年，以及西蒙，1989年），即人们的行为——创造他们自己的历史——不仅受自然世界的约束，也受他们所生于其中的文化和社会的约束。因此，我们的理性在前几代人创造的一套规则中运作，而我们最多只是模糊地意识到这些规则。因此，存在着“理性秩序概念的第二种形式”，即生态理性，

But this account has things backward, putting Descartes before the horse. The horse is history. Smith begins with the insight famously expressed by Marx (though attributed by Smith to Hayek, 1948, and Simon, 1989) that people act—make their own histories—bound not only by the natural world, but also by the cultures and societies into which they are born. Thus, our rationality acts within a set of rules created by the generations preceding us and of which we are at best dimly aware. There is thus a “second of concept of rational order,” ecological rationality,

生态理性是一种未经设计的生态系统，源于文化和生物进化过程：自成体系的行为原则、规范、传统和道德。生态理性运用理性——理性重构——来审视基于经验和民间知识的个体行为，而这些个体在运用建构主义工具进行决策方面能力有限；理解人类文化中涌现的秩序；探索文化和生物遗产中规则、规范和制度所蕴含的潜在智慧，这些智慧源于人类互动，而非人类的刻意设计。人们遵循规则，却无法将其清晰地表达出来，但规则是可以被发现的。

as an undesigned ecological system that emerges out of cultural and biological evolutionary processes: home grown principles of action, norms, traditions, and morality. Ecological rationality uses reason—rational reconstruction—to examine the behavior of individuals based on their experience and folk knowledge, who are ‘naïve’ in their ability to apply constructivist tools to the decisions they make; to understand the emergent order in human cultures; to discover the possible intelligence embodied in the rules, norms and institutions of our cultural and biological heritage that are created from human interactions but not by deliberate human design. People follow rules without being able to articulate them, but they can be discovered.

传统比你更聪明。吉格伦泽和他的团队致力于探索人类通过生物和文化进化习得的原则——快速而节俭的启发法。对他们来说，理性的衡量标准是适应性结果，而非遵循规范主义者的明确规则。

Tradition is smarter than you are. Gigerenzer and his team work to discover the principles—the fast and frugal heuristics—that people have acquired through biological and cultural evolution. For them, rationality is measured by adaptive outcome, not following normativists’ explicit rules.

接下来的两章将探讨认知科学的生物学，首先从认知神经科学开始，这是近因关系的底层。最后一章将探讨终极因果关系，即人类认知过程的演化，并略微涉及遗传学，终极因果关系与近因关系在此交汇。

Our next two chapters are about the biology of cognitive science, beginning with cognitive neuroscience, the bottom level of proximate causation. The last chapter is about ultimate causation, the evolution of human cognitive processes, with a bit of genetics, where ultimate causation meets proximate causation.

外卖

Takeaway

传统比你聪明得多。正如我们将在最后一章所学到的，文化能力是智人的杀手级应用。我们不必依赖达尔文式的自然选择在漫长的世代中缓慢地发挥作用，而是依赖于适应性的洞见和实践，它们可以通过融入文化而立即传承下去。重要的是，智慧的文化编码并非总是有意识的。传统就像具身认知：它们植根于实践，而非理念。例如，杰出的武士刀日本武士的^{15 个}武器都是根据编码尚未发现的冶金知识的传统制作的。

Tradition is much smarter than you are. As we will learn in the final chapter, the capacity for culture is Homo sapiens’ killer app. We don’t have to depend on the slow action of Darwinian selection acting over long generations of change, but on adaptive insights and practices that can be passed on immediately by being built into cultures. Importantly, cultural encoding of wisdom is not always conscious. Traditions are like embodied cognition: They are embedded in practice, not ideas. For example, the outstanding katana of the Japanese samurai¹⁵ were made according to traditions that encoded metallurgical knowledge yet to be discovered.

抛硬币。说真的。莱维特（2021）发现，通过抛硬币暗示做出生活改变（比如寻找新工作等重大改变）的人比对照组更有可能做出改变，并且比没有抛硬币暗示做出改变的人对改变感到更快乐。莱维特的结论是，人们在估计自己能从生活改变中获益多少时过于保守，而抛硬币有助于改善这种偏见。

Flip a coin. Seriously. Levitt (2021) has found that people who made a change in their lives—serious changes such as looking for a new job—suggested by the flip of a coin were more likely to change than controls and happier with their change than those who did it without the flip. Levitt concludes that people are too conservative in estimating how much they might benefit from life changes and the coin flipping helps ameliorate this bias.

用外语思考。Hayakawa 等人（2016）发现，如果人们用成年后学习的外语思考，他们会做出更功利的选择。他们认为，外语比母语更能减少干扰情绪。

Think in a foreign language. Hayakawa et al. (2016) found that people made more utilitarian choices if they thought about them in a foreign language learned in adulthood. They proposed that the foreign language evoked less distracting emotion than native language.

向神经多样性人群学习。Rozenkrantz、D'Mello 和 Gabrieli（2021）回顾了关于自闭症谱系障碍 (ASD) 的研究，结果表明，在许多情况下，ASD 参与者比对照组做出的选择更符合规范理性，包括在我们评测过的经济游戏中，而且他们表现出的认知偏差更少。他们得出结论，ASD 患者比其他人更少受情绪干扰，更少依赖直觉，更容易受到细节的引导。

Learn from the neurodiverse. Rozenkrantz, D’Mello, and Gabrieli (2021) review research on autism spectrum disorder (ASD), showing that, in many cases, ASD participants make more normatively rational choices than controls, including in economic games such as we reviewed, and they show fewer cognitive biases. They conclude that people with ASD are less distracted by emotion, less reliant on intuition, and more guided by details than others.

阅读材料

Readings

保持坚忍。不要“诉诸情感”。以上所有研究结果表明，情绪化会降低信息处理和决策能力，甚至连第一点都做不到——人们过于害怕改变。以掌控情绪（从而运用逻辑）为核心的哲学，正是古老的斯多葛主义。以下是一些书籍。

Be a stoic. Don’t “reach out with your feelings.” All the findings above suggest that emotional involvement can degrade information processing and decision making, even the first one—people are too fearful of change. The philosophy that sets at its center mastering one’s emotions—hence, their involvement with logic—is ancient Stoicism. Here are some books.

罗马最伟大的两位斯多葛学派学者分别是罗马奴隶爱比克泰德（生于公元65年左右）和罗马皇帝马可·奥勒留（公元121-180年）。16爱比克泰德的名言被汇编成一本简短的斯多葛学派手册《手册》^（ https://www.gutenberg.org/files/45109/45109-h/45109-h.htm）。另请参阅英伍德（Inwood）编著的《牛津简明斯多葛学》 （2018年）。奥勒留的斯多葛学派思想在其自传《沉思录》（http://classics.mit.edu/Antoninus/meditations.html）中有所体现。心理学家唐纳德·罗伯逊（Donald Robertson）撰写了两本关于斯多葛主义的著作，分别是《如何像罗马皇帝一样思考：马可·奥勒留的斯多葛哲学》（纽约：圣马丁出版社，2019年）和《认知行为疗法（CBT）的哲学：作为理性与认知心理治疗的斯多葛哲学》（伦敦：劳特利奇出版社，2019年）。将斯多葛主义视为一种伦理哲学，参见 LC Becker 的《新斯多葛主义：修订版》（新泽西州普林斯顿：普林斯顿大学出版社，2017 年）。

Two of the greatest Roman Stoics were Epictetus (born c. 65 CE), a Roman slave, and Marcus Aurelius (121–180 CE), a Roman emperor.¹⁶ Epictetus’s sayings were collected into a brief Stoic handbook, the Enchiridion (https://www.gutenberg.org/files/45109/45109-h/45109-h.htm). See also the Oxford Short Stoicism by Inwood (2018). Aurelius’s Stoicism shows in his autobiography, Meditations (http://classics.mit.edu/Antoninus/meditations.html). Psychologist Donald Robertson has written two books on Stoicism, How to Think Like a Roman Emperor: The Stoic Philosophy of Marcus Aurelius (New York: St. Martin’s, 2019) and The Philosophy of Cognitive-Behavioural Therapy (CBT): Stoic Philosophy as Rational and Cognitive Psychotherapy (London: Routledge, 2019). For a scholarly treatment of Stoicism as an ethical philosophy, see L. C. Becker, A New Stoicism: Revised Edition (Princeton, NJ: Princeton University Press, 2017).

由于近期的政治和外交政策事件，理性话题成为新闻热点。心理学家史蒂芬·平克的两本新书分别是：《理性：它是什么，为什么它似乎稀缺，为什么它重要》（纽约：企鹅出版社，2021）；平克可能是在世的最著名的认知心理学家。此外，还有茱莉亚·加莱夫的《侦察兵心态：为什么有些人看得清楚，而有些人看得不清楚》（纽约：企鹅出版社，2021）。加莱夫主持播客http://rationallyspeakingpodcast.org。经济学家布莱恩·卡普兰的著作是《理性选民的神话：民主国家为何选择糟糕的政策》（修订版）（新泽西州普林斯顿：普林斯顿大学出版社，2011）。他所说的理性是指投票是为了自身利益。选民表达的是他们的态度。许多经济学家不投票，因为他们的投票影响选举结果的可能性太小，不值得花费时间和精力去投票。最后，正如阿富汗和伊拉克战争所表明的那样，所谓的专家往往不善于预测未来。谁能做到？是的。菲利普·泰特洛克多年来一直研究预测能力，发现有些人在这方面非常出色：菲利普·泰特洛克和丹·加德纳合著的《超级预测：预测的艺术与科学》（纽约：皇冠出版社，2015年）。成为专家通常与预测无关——它需要一种特殊的性格，最重要的是，即使会令人尴尬，也要乐于修改自己的预测（做一个坚忍的人）。任何人都可以注册他的“良好判断力项目”，https://www.gjopen.com。

Because of recent political and foreign policy events, the topic of rationality is much in the news. Two new books by psychologists are Steven Pinker, Rationality: What It Is, Why It Seems Scarce, Why It Matters (New York: Penguin, 2021); Pinker is probably the most famous living cognitive psychologist. Also, Julia Galef, The Scout Mindset: Why Some People See Things Clearly and Others Don’t (New York: Penguin, 2021). Galef hosts a podcast http://rationallyspeakingpodcast.org. Economist Brian Caplan has written The Myth of the Rational Voter: Why Democracies Choose Bad Policies, rev. ed. (Princeton, NJ: Princeton University Press, 2011). By rational he means voting in one’s self-interest. Instead, voters express their attitudes. Many economists don’t vote because the chance that their vote swings an election is too small to be worth the time and effort it takes to vote. Finally, as the wars in Afghanistan and Iraq showed, so-called experts are often terrible at predicting the future. Can anyone? Yes. Philip Tetlock has studied forecasting ability for many years and has discovered that some people are great at it: Philip Tetlock and Dan Gardner, Superforecasting: The Art and Science of Prediction (New York: Crown, 2015). Being an expert often has nothing to do with it—it takes a particular sort of personality, most importantly, being open to revising one’s forecasts even if it’s embarrassing (be a Stoic). Anyone can sign up for his Good Judgment Project, https://www.gjopen.com.